[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7ATVX5RRBGHT90PZ0P54ZCX":3,"skill-same:01M7ATVX5RRBGHT90PZ0P54ZCX":53},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":25,"commitSha":26,"skillDirectory":27,"forksCount":14,"repositoryHtmlUrl":28,"repositoryHomepageUrl":6,"ownerHtmlUrl":29,"ownerBlogUrl":30,"ownerDescription":31,"updatedAt":32,"install":33,"resources":46},"01M7ATVX5RRBGHT90PZ0P54ZCX","specrew-design-workshop",null,"Run Specrew's per-lens design workshop and collaborative design-analysis. Use whenever you work the design lenses for a feature: at specify\u002Fintake (the lens workshop) and at the design-analysis stop (co-design the architecture), and RE-INVOKE at the start of EACH new lens (architecture, data, ui-ux, security, integration, devops, requirements\u002FNFR, observability, component). Triggers: design, design lens, lens workshop, design-analysis, architecture, trade-offs, co-design, explore options, decompose, or moving from one lens to the next. Tells you to facilitate each lens as a discussion, surface diagrams the human can actually SEE (console ASCII inline; mermaid\u002Fhtml to a file with a clickable file:\u002F\u002F\u002F link), co-design components\u002Fresponsibilities\u002Fflows WITH the human instead of handing over finished options, capture the agreements, and which per-lens md to load.","alonf\u002FKeyContextAI","alonf","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F554150?v=4","01M7ATEPSCNKG13BVTKA74VG59","Alon Fliess","01M7ATEPSZ4YMBM9W25SKM3GES",0,"MIT","https:\u002F\u002Fgithub.com\u002Falonf\u002FKeyContextAI\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fspecrew-design-workshop\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: \"specrew-design-workshop\"\ndescription: \"Run Specrew's per-lens design workshop and collaborative design-analysis. Use whenever you work the design lenses for a feature: at specify\u002Fintake (the lens workshop) and at the design-analysis stop (co-design the architecture), and RE-INVOKE at the start of EACH new lens (architecture, data, ui-ux, security, integration, devops, requirements\u002FNFR, observability, component). Triggers: design, design lens, lens workshop, design-analysis, architecture, trade-offs, co-design, explore options, decompose, or moving from one lens to the next. Tells you to facilitate each lens as a discussion, surface diagrams the human can actually SEE (console ASCII inline; mermaid\u002Fhtml to a file with a clickable file:\u002F\u002F\u002F link), co-design components\u002Fresponsibilities\u002Fflows WITH the human instead of handing over finished options, capture the agreements, and which per-lens md to load.\"\ndisallowed-tools: AskUserQuestion\n---\n\n# specrew-design-workshop\n\n**Type**: Facilitation Skill\n**Schema**: v1\n**Status**: Active design-lifecycle conduct\n\n## Purpose\n\nDrive the design lenses as a real, point-of-use **workshop** — not a checklist skimmed once at launch. Each\ntime you reach a lens, this skill is loaded fresh so the conduct is in focus exactly when you use it. **It is\nself-contained: one load gives you everything you need for the current lens.** Because most hosts do not\nguarantee a fresh reload per turn, **re-invoke this skill (or re-read it) at the start of each new lens.**\n\n## The Big Picture (read this first, every time)\n\nYou are facilitating a design conversation with a human, **one lens at a time**. The lenses live in\n`extensions\u002Fspecrew-speckit\u002Fknowledge\u002Fdesign-lenses\u002F` (deployed under the project's design-lens catalog):\n`architecture-core`, `component-design`, `requirements-nfr`, `ui-ux`, `data-storage`, `security-compliance`,\n`integration-api`, `devops-operations`, `observability-resilience`, `code-implementation`. For each lens you work:\n\n1. **Load that lens's md** (`design-lenses\u002F\u003Clens-id>.md`) for its `## Design Decision Points` and\n   `## Workshop Conduct` — that is the lens's focused agenda. Do not improvise it from memory.\n2. **Facilitate** the discussion for that lens (the method below).\n3. **Record** the decisions + agreement.\n4. **Re-invoke this skill** and go to the next lens.\n\nThe per-lens *knowledge* is in the lens md; this skill is the *method* that ties the lenses together. Keep\nboth in view.\n\n**Render before you ask — the question may only reference what is on screen (A6\u002FA7\u002FFR-037).** Before you\nraise ANY confirm\u002Fapprove question (as a typed \"does this work? \u002F move on? \u002F approve\" prompt) about the lens\n**agenda + depths**, a per-lens diagram, the\ncomponent map, an options\u002Ftrade-off set, or a design verdict — the thing you are asking about MUST already be\n**rendered in your assistant message in THIS exchange**, in prose \u002F console-ASCII the human can see. The\nquestion may reference ONLY content that is on screen. **Never** ask from a count (\"8 lenses\", \"13\ncomponents\"), a summary (\"the agenda\", \"the map\"), or a \"shown above \u002F as proposed\" that was not actually\nshown. The failure mode this exists to stop is **menu-before-render**: render + explain, THEN ask. **Be verbose**\n— explain the agenda \u002F diagram \u002F map \u002F options fully in prose first, and make the question and choices\nself-explanatory (not a terse \"does this work? — 8 lenses\").\n\n**Typed-turn authority rule — every host:** workshop questions are visible prose and the human answers by typing.\nOffer known choices as a **numbered prose list answered by typing** (or ask concise free-form prose); do not call a\nstructured question\u002Fmenu tool anywhere in the product-domain phase, agenda confirmation, or lens work. A picker\ncan swallow rendered context. **Ctrl+O is no answer**: Copilot reports it as `User skipped question`, which a model\nthen misread as delegation. Neither dismissal nor a picker selection is workshop authority. If one is invoked anyway,\nre-render the unanswered question as prose and wait: never infer defaults, `human-delegated`, or `human-skipped`.\nClaude additionally removes `AskUserQuestion` at capability level. The other hosts are held by the hook-owned\ntyped-turn receipt: no typed reply means no receipt and the controller\u002Fboundary refuses the claimed agreement.\nAfter each typed reply, read the latest matching line from `.specrew\u002Fruntime\u002Fworkshop-authority.jsonl` and copy\nits `receipt_id` into the record you are about to write. If there is no matching receipt, re-render the question\nand stop; never manufacture an id or edit the hook-owned store.\n\n**The `file:\u002F\u002F\u002F` links go in your prose before the question, too (dogfood finding).** When a confirm\u002Fapprove\nquestion references or asks the human to review an artifact — the spec, a lens workshop record, the\ndesign-analysis, a diagram file — emit the **bare clickable `file:\u002F\u002F\u002F` links in your assistant message in THIS\nexchange**, never only inside structured question fields and never as a bare \"see the file above\". Structured\nquestion UIs may not linkify `file:\u002F\u002F\u002F`; Claude's picker is disabled for this skill because it can drop the\nwhole preceding message. Rule 14A's clickable-`file:\u002F\u002F\u002F` guarantee must therefore hold *at* the confirmation,\ncarried by visible prose.\n\n**(A8\u002FFR-041) Open each lens with a presentation + an open question, never a menu** (The Method step 3) — this is\nwhat actually rendered the lens content on Claude in the dogfood (the per-lens open has no competing menu, so\nthe content cannot collapse into one). **Governing model (dogfood-proven):** content whose next move is *open\ndiscussion* renders reliably. On Claude, content immediately before `AskUserQuestion` can be swallowed by the\npicker; the deployed skill therefore removes that tool for the whole workshop. This is a capability-level\nguard, not another conduct instruction or hidden marker. An earlier catalog-at-open front-load was reverted\nbecause it still skimmed on Claude and was redundant on prose hosts that render the agenda inline.\n\n## First stage — the product-domain phase (run before everything, Feature 176)\n\nBefore the lens applicability agenda (The Method below), run the **product-domain** first-stage\nphase — the pre-technical product\u002Fproblem grounding. It is **always applicable** and runs FIRST,\nbefore any technical lens or the applicability questionnaire. It is NOT a row in\n`applicability-map.json`; it is the first-stage phase ahead of the deterministic selector.\n\n0. **Establish durable pre-agenda controller state before the first product-domain question.** Immediately\n   after the feature scaffold creates its nonempty `spec.md`, and only when the feature-level\n   `lens-applicability.json` is absent, run:\n   `pwsh -NoProfile -File .specify\u002Fextensions\u002Fspecrew-speckit\u002Fscripts\u002Finitialize-workshop-controller-state.ps1 -ProjectRoot . -FeatureRef \u003Cexact-feature-ref>`.\n   The helper atomically writes the strict feature-only `agenda_status: pending-confirmation` shape with empty\n   `selected` and `workshop` collections. It refuses to overwrite an existing artifact. On resume, read and\n   honor the existing exact-scoped artifact instead of rerunning the helper. This is controller truth for the\n   product-domain and agenda-confirmation questions; it is not a hidden model marker, environment variable, or\n   transcript heuristic.\n1. **Load the lens** `design-lenses\u002Fproduct-domain.md` for its decision areas, depth model,\n   evidence vocabulary, run cadence, and conduct.\n2. **Select the depth** (Light \u002F Standard \u002F Deep) by **risk and novelty**, and say why. A tiny\n   utility gets a Light pass that records why deeper discovery is not warranted.\n3. **Reframe a solution-first request into the problem.** When the human asks to \"build X\", surface\n   who it is for, the pain\u002Fjob, the MVP, the non-goals, and the constraints BEFORE any design. If the\n   requested feature is not aligned to the pain \u002F MVP \u002F constraints \u002F alternatives, surface that\n   before plan.\n4. **Capture the product context** at the selected depth (users and stakeholders; pain\u002Fjob and\n   current workaround; existing system\u002Fcontext; constraints; outcomes; MVP\u002Fnon-goals\u002Fvision;\n   alternatives at Standard\u002FDeep), and **tag every material statement** with its evidence quality\n   (`known` \u002F `assumed` \u002F `unknown` \u002F `research-needed`; a `research-needed` carries `load_bearing`).\n   Honest tags prevent confident product fiction.\n5. **No batch confirmation (FR-009).** The phase cannot be satisfied by approving an agenda or a batch\n   \"confirm all\". It needs scoped product-domain confirmation, or an explicit, honestly recorded\n   \"you decide\" \u002F \"skip\". Record the provenance (`human-confirmed` \u002F `human-delegated` \u002F\n   `human-skipped`) and its matching scope.\n6. **Persist both records** under `specs\u002F\u003Cfeature>\u002Fworkshop\u002F`: `product-domain.yml` (structured,\n   schema-validated against `contracts\u002Fproduct-domain.schema.json`) and `product-domain.md`\n   (human-readable), and summarize the decisions into `spec.md` — not as the sole source. The\n   specify-gate floor REQUIRES this for a substantive feature before specify syncs; an absent record\n   is surfaced, never silently skipped.\n\n**Run cadence**: the product-domain phase runs before EVERY feature at adaptive depth — not once. In\nV1 every feature is `context_scope: feature_standalone`; once Proposal 162 ships, later features run\nin delta mode (`feature_delta`) against the inherited product baseline (`product_baseline`).\n\nOnly after the product-domain phase is captured do you move to the lens applicability agenda below.\n\n## The code-implementation lens (auto-on for code features; Feature 177)\n\nThe `code-implementation` lens captures *how the code is written* (implementation craft) as binding\nconstraints for implement. It is **always-applicable for any feature that writes code** (auto-on; skip\nonly doc-only \u002F config-only slices, with a recorded reason). It is **conduct-driven** — NOT a row in\n`applicability-map.json` (drift D-001), exactly like `product-domain`: you include it for code features\nwithout a yes\u002Fno applicability question. Work it **after the technical lenses** (it depends on the\nresolved stack + the architecture decision). Load its md\n`design-lenses\u002Fcode-implementation.md` for the decision spine, per-stack dilemmas, the grouping model, the\nrun-cadence, and the full conduct. The conduct in brief:\n\n1. **Source of code-rules truth FIRST** (the Figma-equivalent question): ask whether the human has an\n   existing coding guideline OR one or more **example projects** to emulate (a GitHub repo, a local path,\n   or other) for code style, language constructs, and patterns — or none.\n2. **Assisted ingestion** (when a guideline \u002F example project is provided; agent-reasoning, no parser):\n   map it onto the `code-rules.yml` catalog (auto-check matches, flag conflicts for the human), and extract\n   non-catalog conventions as custom rules with provenance (`from-guideline` \u002F `from-example-project` +\n   the source ref). Company\u002Forg-level rules persist to the reusable project overlay `code-rules.local.yml`\n   (additive + per-rule override; never drops a shipped rule).\n3. **Resolve the stack**, then present the **grouped, pre-checked set\u002Funset checklist** — baseline stated\n   as a summary (exceptions only), the consequential decision-prompts paced (offer all-at-once OR\n   one-at-a-time), applicability-filtered rules shown only when their context applies. **Never a flat wall.**\n   **\"Use the defaults\" means one coherent bundle, not every familiar mechanism at once.** Check that the\n   chosen defaults can coexist before recording them. For example, a small manually composed .NET CLI may own\n   one shared `HttpClient`; choosing `IHttpClientFactory` implies factory\u002Fservice-provider composition and must\n   not be paired with \"manual `new()`, no DI container\". Resolve contradictions with the human in this lens.\n4. **Dependency selection (FR-013)**: present **\"use existing project tools \u002F no new dependency\" first**\n   plus options; for any chosen dependency capture version, license, source org, canonical URL, maintenance\n   signal, security\u002Fadvisory status, compatibility, cost\u002Fquota, coupling weight, replaceability, and test\n   implications into the manifest `dependency_policy`.\n5. **Capture** the selections\u002Fdecisions\u002Fcustom-rules\u002Fdependency-policy by authoring the per-feature\n   `implementation-rules.yml` manifest (reference-by-ID), schema-validated against the manifest schema\n   `implementation-rules.schema.json` (beside the lens in the design-lens catalog), plus the human-readable\n   `workshop\u002Fcode-implementation.md`; record the lens in `lens-applicability.json`. (Like `product-domain`,\n   you author the record by hand following the schema -- you do NOT call a PowerShell writer; the lens\n   helper validates, it is not invoked mid-workshop.) Reviewer setup is part of this capture: list the live\n   hosts, present the marked default, record one concrete host + authorization reference, and run the explicit\n   `specrew review --host ... --authorization-ref ...` setup command. Never leave a completed code lens with\n   an unresolved `auto-select`; that state cannot launch a reviewer in a fresh project.\n6. **Run cadence**: the rules are mostly product-level — decide once at a product-level workshop, inherit\n   per feature, re-open only the parts a new technology or programming language changes\n   (`context_scope` hooks; V1 `feature_standalone`; forward-compatible with Proposal 162).\n\nThe **implement-time half** is the separate `specrew-code-rules` skill, which reads this manifest while the\ncoding agent writes code and surfaces the rules task-scoped. The acceptance gate is the deployed dogfood\n(SC-004\u002FSC-007\u002FSC-008), not unit-green.\n\n## The Method (the same for every lens)\n\n1. **Frame the phases + hand over the agenda (A6\u002FFR-034, A7\u002FFR-040).** Tell the human up front: the workshop\n   *gathers* inputs and constraints; the system **structure** (components, responsibilities, flows) is designed\n   *with* them at the design-analysis step — not decided in intake. So they know the collaboration is coming.\n   **Before the heavy prep, keep them oriented (FR-040):** say plainly that you are *preparing the workshop and\n   it will take a moment* (so a pause does not read as a hang), and hand them the **agenda as an assignment** —\n   list the lenses you will work and, for each, the decision it will ask of them — so they can think or research\n   while you load. The wait becomes preparation, and a prepared human engages per-lens (which is what keeps the\n   integrity rule in step 6 honest). **Tell the human they can just talk (dogfood finding):** in the same\n   framing, say plainly that at any lens, if a question is unclear, they want to open a file, or they need more\n   detail, they can simply *type* it (for example \"explain more\") instead of picking a menu option — you will\n   explain, then re-ask. Typed\n   conversation is the required path on every host; do not use the host's structured picker during the workshop.\n   **Product-domain owns product facts, not technical decisions.** The Crew proposes the product-domain depth\n   from risk and novelty and the human confirms or adjusts that depth in their typed product-domain reply; do not persist\n   the selected product-domain depth before that typed human confirmation. Product-domain may ask whether a\n   technology is already externally fixed as a constraint. It must not ask the human to choose runtime\u002Fstack,\n   output format, timeout\u002Freachability semantics, concurrency, or dependencies: those belong to their selected\n   technical lens, where each receives its own turn and typed receipt.\n2. **Infer applicability, then confirm (A4\u002FFR-025).** Propose which lenses apply WITH your reasoning; ask the\n   human only to confirm or adjust. Never make them answer obvious yes\u002Fno applicability; never silently\n   auto-resolve a material area. **Render the agenda IN-BAND before asking for confirmation — fill this\n   template; do NOT cram the lens list into a question UI** (a prose \"render first\" gets skimmed on some hosts;\n   the filled template is what the human reads while you prepare):\n\n   ```text\n   Workshop agenda — \u003CN> lenses\n\n   \u003Clens-id> (\u003Cfull | medium | light>) — \u003Cthe decision this lens will ask you to make for THIS feature>\n   \u003Clens-id> (\u003Cfull | medium | light>) — \u003Cthe decision …>\n   ...\n   Skipped: \u003Clens-id> — \u003Cwhy it does not apply here>\n   ```\n\n   Fill ONE line per applicable lens with its depth and the **concrete decision it raises** (not just the lens\n   name); render the whole filled block in your message, THEN ask the human to confirm or adjust it as visible\n   prose. Wait for a typed response on every host.\n   **Show the negative selection too.** One `Skipped:` line is required for **every** technical lens that is not\n   selected, with a feature-specific reason. Then ask, in plain language, whether the human wants to confirm or\n   change the selection. The first lens must not open on the same turn as the agenda: typed human confirmation is\n   the transition. A count such as \"5 lenses\" is not reviewable unless the omitted set is visible too.\n   **Render the authority-bound agenda before asking.** Build one JSON object shaped as\n   `{\"selected\":[{\"lens\":\"architecture-core\",\"depth\":\"medium\",\"decision\":\"...\"}],\"skipped\":[{\"lens\":\"ui-ux\",\"reason\":\"...\"}]}`,\n   then invoke `& .specify\u002Fextensions\u002Fspecrew-speckit\u002Fscripts\u002Fconfirm-workshop-agenda.ps1 -ProjectRoot . -FeatureRef \u003Cfeature> -AgendaJson $agendaJson -RenderOnly`.\n   Send the command's complete output as one unchanged, contiguous block. A short introduction or closing is\n   allowed, but may not replace, paraphrase, or split the canonical block. The hook binds the typed reply to the\n   structured agenda digest after proving that complete block was visible. Wait for the human's\n   typed response. **The moment they confirm, PERSIST it before opening lens 1** by invoking the same\n   command with the identical `$agendaJson` and without `-RenderOnly`. If the selected set, depths, decisions,\n   skipped set, or reasons change, render the new canonical agenda and ask again; an older receipt cannot authorize it.\n   The writer changes `agenda_status` to `confirmed` only when selected + skipped cover the whole deployed\n   technical-lens catalog and records `agenda_confirmation: human-confirmed` with\n   `agenda_confirmation_scope: lens-selection`. It also requires hook-owned typed reply receipts for product-domain\n   and the complete agenda; Ctrl+O or a picker result produces neither. It refuses incomplete coverage or missing\n   authority. The per-lens `workshop` records\n   are added later, as each lens completes per step 6. A resume can only compute the remaining agenda if the agenda\n   itself is on disk; an agenda that lives only in scrollback is lost on exit (observed: an unpersisted agenda made\n   a resuming host re-run specify instead of continuing the workshop).\n   If the named action does not clear a refusal on the next attempt, stop; do not retry again and do not write\n   workshop records by hand to get past it. Report calmly in the user's project terms: what you were doing; that\n   their answers are safe and nothing has been lost (only when that is true); what you could not complete without\n   diagnosing or assigning fault; one concrete next action you can perform; and ask the human for their approval.\n   Keep machinery diagnosis in the drift record, not in the message the human must act on.\n   When the human approves preparing a repair for inconsistent pre-agenda progress, use the governed recovery\n   surface instead of editing `lens-applicability.json`: run\n   `pwsh -NoProfile -File .specify\u002Fextensions\u002Fspecrew-speckit\u002Fscripts\u002Frepair-workshop-controller-state.ps1 -ProjectRoot . -FeatureRef \u003Cexact-feature-ref> -Request`,\n   render its complete proposal, and wait for the exact typed authorization it names. Only after that reply has\n   been captured may you run the same command with `-Apply`. The repair is state-hash-bound, preserves the saved\n   product-domain records, writes an audit record, and refuses if anything changed after authorization.\n   **Agenda confirmation is not lens-question confirmation.** This confirm point approves only the selected\n   lens list and depths. It does NOT answer the lenses. Do NOT offer or accept a batch shortcut such as \"Confirm\n   all as proposed\", \"approve all lens decisions\", or \"use the proposed decisions for every lens\" as\n   `human-confirmed` \u002F `lens-question`. After the agenda is confirmed, every selected lens still needs its own\n   lens-specific turn and its own human answer, explicit delegation, or explicit skip.\n3. **Per-lens facilitated discussion — open with a presentation + an OPEN question, never a menu first (A4\u002FFR-025, A8\u002FFR-041b).**\n   The **first turn of every lens** MUST be you *presenting* the lens — its decision points, and as it develops\n   its diagram \u002F component map — followed by an **open, free-text question** (\"how should we approach this?\",\n   \"what are your constraints?\"). Do **NOT** open a lens with an `AskUserQuestion` \u002F structured menu: that is the\n   move that lets the content collapse into the menu's question field and never get rendered (the A8\n   `AskUserQuestion` tool-gravity failure). On Claude the tool is unavailable for the entire skill; use visible\n   prose and typed numbered choices. On every other host the same choices remain visible prose with a typed reply;\n   do not switch to a structured menu after rendering. Binary\n   test: did this lens open with a rendered presentation, or with a menu? Open with the presentation.\n   **One selected lens = one lens turn.** Do NOT bundle several selected lenses into one combined presentation\n   and one \"confirm all\" question, even for a tiny feature or light-depth lenses. A lens turn may summarize the\n   already-confirmed agenda, but it must focus on exactly one lens's decision points, ask for that lens's answer,\n   and wait for the human (or an explicit \"you decide for this lens\" \u002F \"skip this lens\") before moving on.\n   This is also a question-content boundary: a product-domain turn must not harvest runtime\u002Fstack, output,\n   timeout\u002Freachability, concurrency, or dependency decisions owned by later technical lenses. A receipt for\n   one lens cannot authorize another lens's answer.\n   **Keep the controller-owned workshop state durable and complete (FR-055\u002FFR-056).** Before you stop and wait\n   for an answer, ensure the applicability artifact for the CURRENT workshop scope exists and reflects either\n   the strict feature-level pre-agenda state or the current confirmed agenda. During specify\u002Fintake, authority is\n   the feature-level\n   `specs\u002F\u003Cfeature>\u002Flens-applicability.json`; during design analysis it is the exact iteration's\n   `specs\u002F\u003Cfeature>\u002Fiterations\u002F\u003CNNN>\u002Flens-applicability.json`. Never invent an iteration during feature intake.\n   Do not rely on a model-authored hidden marker, an environment variable, or a host question-tool transcript:\n   hosts can omit, transform, or swallow those surfaces. The Stop controller derives `active` only from the exact\n   scoped artifact: either its strict feature-only `pending-confirmation` shape with a valid scaffolded spec, or\n   its confirmed nonempty, unique selected agenda and strict ordered completion records. While that state is\n   valid and incomplete, ordinary product-domain, agenda, and lens questions remain conversational and the\n   generic five-section non-boundary packet is suppressed; a real lifecycle boundary still has precedence.\n   **Finish each lens durably in this order:** write the nonempty `workshop\u002F\u003Clens-id>.md` decision record first,\n   then persist that lens's full step-6 entry with `moved_on: true`. On the final selected lens, that second write\n   makes the workshop `complete` and ordinary Stop behavior resumes immediately. A loose flag, missing record,\n   duplicate\u002Fout-of-order lens, malformed artifact, or mismatched confirmation is invalid and cannot keep the\n   exception active. Never render the generic five-section packet merely to ask or answer an ordinary lens\n   question; use the six-section packet only for a genuine lifecycle boundary.\n   **Carry binding decisions across lenses.** Before opening a new lens, reread the completed records' durable\n   `bindings`. A delegated\u002Fdefault decision may fill an unresolved gap, but it MUST NOT contradict an earlier\n   human-confirmed binding. If the human intentionally changes a binding, reconcile every affected lens record\n   to the same value before moving on; never leave two durable answers for one architectural fact.\n   **Pace a dense lens — after presenting, you MUST offer all-at-once OR one-at-a-time (A8\u002FFR-041b UX, every host).**\n   A lens with several decision points (architecture-core, component-design, security-compliance) lands as an\n   overwhelming **wall** if you present everything and end with one open question that secretly bundles five\n   subjects — and a numbered list of subjects reads like pickable options when it is not. So on a dense lens\n   (**three or more** decision points) your closing move **MUST** offer the human a **pacing choice** — *answer it\n   all in one go, or have me take you through the decisions one at a time* (then ask decision 1, wait for the\n   answer, then decision 2, …, each as its own focused question). This is **not optional on a dense lens and\n   applies on every host** — a single open question covering five subjects is as hard on Copilot\u002FCodex\u002FAntigravity\n   as on Claude; do not bundle. Respect their pick; never force the whole wall on someone who wants to step\n   through it. A light single-decision lens skips the offer — just ask the one open question. Then, for\n   the current lens, raise its decision points (from its\n   md), offer options where useful, capture the human's needs + decisions + explicit agreement, and **iterate\n   until the human says \"move on\"** before the next lens. Adapt depth to the user-profile expertise dials\n   (concise where high; explain + recommend a default where low). Right-size — not a fixed nine-lens marathon.\n   **Match the question FORM to the question**: for a discrete, enumerable choice (e.g. decomposition\n   vocabulary — IDesign \u002F Clean Architecture \u002F modular; one service vs split; fixed vs open taxonomy), spell\n   out every option and an explicit \"other \u002F let me explain\" path so the human can pick fast. This is a numbered\n   prose list answered by typing on every host. For a genuinely open question, discuss in prose. Both are fine — do not\n   force a discrete pick into long prose, nor an open design question into a rigid one-shot MCQ. **Surface EVERY\n   selected lens to the human and get a real answer before you record it (A7\u002FFR-038):** intake is NOT \"specific\n   enough\" until each selected lens has either the human's confirmation OR an explicit \"you decide \u002F skip\" from\n   them. You may NOT decide after a few questions that intake is done and then write up the remaining lenses\n   yourself — that is the exact failure this rule exists to stop. When you move to the next lens (loading it\n   lazily), announce it so the pause is legible: *\"preparing lens X of N: &lt;lens&gt; — get ready, this one\n   decides …\"* (FR-040).\n4. **Surface visuals IN-BAND so the human can SEE them (A5\u002FA6\u002FFR-030–FR-031\u002FFR-037).** On a terminal\u002Fconsole\n   host a fenced ```mermaid``` block is **source text, not a rendered picture** — only **console ASCII**\n   actually renders inline. So the diagram you show the human MUST be **console ASCII art rendered directly in\n   your message** (boxes, arrows, labels) — reach for ASCII first and by default. Use mermaid \u002F svg \u002F html only\n   as a *richer, additional* artifact written to a file under the workshop folder — and when you do, you MUST\n   emit its clickable `file:\u002F\u002F\u002F` link in the SAME message. A diagram the human cannot see is NOT surfaced;\n   never let the human reach a decision while a diagram lives only in a file or only as mermaid source. A\n   per-lens diagram is **expected** for the structural lenses (architecture, data, security, integration,\n   devops) and any UI-bearing feature (ui-ux: a layout sketch). Intake is bidirectional: offer to plot the\n   lens's diagram AND ask whether the human has an existing diagram\u002FFigma\u002Fwhiteboard photo to bring. **At ANY\n   approval point — \"approve\" \u002F \"move on\" \u002F \"co-design\" \u002F \"does this work as the baseline\" — the diagram or\n   component map MUST be rendered in console-ASCII IN THE SAME MESSAGE as the question; NEVER stand in a\n   reference to it (a file path, or a bare count like \"13 components\" \u002F \"4 managers\"). The human approves what\n   is on screen, not what is in a file.** Persist the keeper to the workshop file AFTER and IN ADDITION TO the\n   in-band render (step 6) — never INSTEAD of it. (Empirically: testLenses8 on the Claude host wrote the\n   component map to the workshop file and asked the human to approve \"13 named components\" it never showed,\n   while Copilot and Antigravity rendered it in-band first — writing the file is not surfacing.)\n5. **Co-design — do NOT hand down finished options (A6\u002FFR-035\u002FFR-036).** At the design-analysis stop:\n   - **Co-decide the design method \u002F decomposition style** (DDD bounded-contexts, IDesign volatility-based,\n     modular monolith, microservices, layered) as an expertise-adapted discussion; record the choice as a\n     binding constraint. Not a bare multiple-choice prompt; do not silently assume it.\n   - **Co-build the component map — render the FULL form, never a summary or a count.** Whenever you present\n     the component map (at the component-design lens AND at the design-analysis stop), put this IN YOUR MESSAGE,\n     in order: **(1)** a **console-ASCII diagram** of all components on their layers with dependency arrows;\n     **(2) then a named list grouped by the decomposition vocabulary** you bound at the design-method step —\n     Managers \u002F Engines \u002F ResourceAccessors for IDesign; bounded contexts \u002F aggregates \u002F entities for DDD;\n     layers for layered; services for microservices — with **EVERY component named and its one-line\n     responsibility** (e.g. `TrayClient — owns the tray UI + global hotkey`). **NEVER a bare count** (\"6\n     resource accessors\", \"3 engines\") and never a \"map above\" \u002F file reference — list every component by name.\n     ONLY after the diagram + the full named list are on screen do you ask the human to approve \u002F rename \u002F\n     split \u002F merge \u002F reassign, and walk at least one key **flow** (user + system actions) through it together.\n     **If they ask for a change, re-render the updated diagram + list, then ask again** — iterate until the\n     human agrees the decomposition, responsibilities, and flows are right. **Use this fill-in template so the\n     form is consistent** — copy it into your message and replace every `\u003C...>`, list EVERY component (never a\n     count), group by your chosen vocabulary, render it in-band BEFORE the approve\u002Fchange question, and\n     re-render the filled template on any change:\n\n     ```text\n     Proposed component map\n\n     \u003Cconsole-ASCII diagram: every component on its layer, dependency arrows pointing inward>\n\n     \u003Cvocabulary group 1 — Managers | Bounded Contexts | Layers | Services>:\n       \u003CComponentName> — \u003Cone-line responsibility>\n       \u003CComponentName> — \u003Cone-line responsibility>\n     \u003Cvocabulary group 2 — Engines | Aggregates | ...>:\n       \u003CComponentName> — \u003Cone-line responsibility>\n     \u003Cvocabulary group 3 — Resource Accessors | Repositories | ...>:\n       \u003CComponentName> — \u003Cone-line responsibility>\n\n     Key flow: \u003Cactor action> -> \u003CComponent> -> \u003CComponent> -> \u003Coutcome>\n     ```\n\n   - **Only then present the remaining trade-off options** (transport, store technology, granularity) for the\n     human's verdict — the leftover consequential choices, not the whole design handed down.\n   - Where the human's architecture expertise is low you MAY drive the decomposition and explain more, but you\n     MUST still confirm the responsibilities and flows with them, not author them silently.\n6. **Capture the agreements (A4\u002FA6\u002FSC-021\u002FSC-025).**\n   - Per-lens workshop record in the feature-level `lens-applicability.json` — set `workshop_intake: true`,\n     `confirmation_required: true`, `agenda_status: confirmed`, the `selected` lens-id list, and a SINGLE\n     top-level `workshop` object\n     **keyed by lens id**, each value carrying the EXACT fields the gate checks: `agenda` (array of questions\n     raised), `decision` (a SINGLE STRING summarizing the decision + agreement), `depth`, `moved_on: true`, and\n     **`confirmation`** — the provenance, one of `human-confirmed | human-delegated | human-skipped` (A7\u002FFR-039,\n     SC-026), plus **`confirmation_scope`** — `lens-question` for `human-confirmed`, `explicit-delegation` for\n     `human-delegated`, or `explicit-skip` for `human-skipped`; and **`human_turn_receipt`**, the receipt id from\n     the latest matching line in `.specrew\u002Fruntime\u002Fworkshop-authority.jsonl`. No receipt means re-render and wait.\n     Exact shape — get it right the first time:\n\n     Record every load-bearing or cross-lens decision in an optional `bindings` object using stable lowercase\n     keys and token values. Reuse the same key whenever another lens touches that decision. Repeated values must\n     match exactly; the controller stops on a conflict before the workshop can finish.\n\n      ```json\n      { \"workshop_intake\": true, \"confirmation_required\": true, \"agenda_status\": \"confirmed\", \"selected\": [\"architecture-core\"],\n        \"workshop\": { \"architecture-core\": { \"agenda\": [\"q1\",\"q2\"], \"decision\": \"what was decided + agreed\", \"depth\": \"full\", \"moved_on\": true, \"confirmation\": \"human-confirmed\", \"confirmation_scope\": \"lens-question\", \"human_turn_receipt\": \"\u003Cmatching receipt_id>\", \"bindings\": { \"article-initiation\": \"on-demand\" } } } }\n      ```\n\n     It is `workshop` -> `\u003Clens-id>` -> fields (NOT `\u003Clens-id>` -> `workshop`), and `decision` is a singular\n     string, NOT a `decisions` array — the inverted nesting or a `decisions` array FAILS the SC-021 gate. Extra\n     fields (such as the `diagram` reference below) are fine. **Persist each keeper diagram you render for a lens** (the trust-boundary diagram, the ERD, the\n     component map, the sequence, etc.) — write the ASCII (or mermaid) to a workshop folder,\n     `specs\u002F\u003Cfeature>\u002Fworkshop\u002F\u003Clens-id>.md`, and set that lens's `diagram` field to a **reference to that\n     file** (the path + a one-line caption), NOT a prose description. A diagram that lives only in the chat\n     scrollback is lost; the workshop folder makes the design reviewable from the artifacts.\n   - **Integrity — never manufacture agreement (A7\u002FFR-038, SC-026).** Set `confirmation: human-confirmed` and\n     `confirmation_scope: lens-question` ONLY for a lens whose substantive workshop questions you actually\n     surfaced and the human confirmed. Lens approval is not workshop-question approval. If the human explicitly\n     said \"you decide\" or \"skip\" for that lens's questions, set `human-delegated` + `explicit-delegation` or\n     `human-skipped` + `explicit-skip` and say so honestly in `decision` — do NOT write a fabricated \"the human\n     agreed to X\" for a lens they never saw. **A batch confirmation is not valid provenance:** if the human says\n     \"confirm all as proposed\" at the agenda or summary level, that confirms only the agenda\u002Fscope and cannot\n     produce `human-confirmed` \u002F `lens-question` for any lens. If the human explicitly delegates a shortcut such\n     as \"you decide the remaining lenses\", record each affected lens as `human-delegated` +\n     `explicit-delegation`, not `human-confirmed`. **Count self-check before you record:** you are about to write\n     N lens records — you must have asked, or been explicitly told to decide\u002Fskip, N times, one lens at a time.\n     A dismissed picker is an absence, never explicit delegation. If you stopped early, go back and surface the rest; you may not declare intake \"specific enough\" and fill in\n     the remaining lenses yourself. The SC-026 gate blocks the specify boundary until every selected lens carries\n     a `confirmation` and matching `confirmation_scope` — but it cannot see transcript truthfulness; that\n     integrity is on you, and the Squad re-dogfood checks it.\n   - At design-analysis, a `## Co-Design Record` in `design-analysis.md` with the agreed\n     component-to-responsibility map + at least one agreed flow + a human-agreed marker, and — when ui-ux is in\n     scope — the **agreed UI\u002Fscreen layout** (the ASCII sketch the human approved). Set `co_design: true` in the\n     iteration's `lens-applicability.json` so the deterministic co-design-record floor applies. An agreement\n     that lives only in the chat scrollback is lost.\n7. **Checkpoint this lens durable, THEN re-invoke for the next so a mid-workshop exit\u002Fswitch is recoverable.**\n   The workshop is long, and an exit or host-switch mid-workshop is expected, not exceptional — so make each\n   lens durable the moment you finish it, never \"all at the end\". BEFORE you move to the next lens:\n   **(a) persist the Markdown decision record FIRST** at the exact current scope's\n   `workshop\u002F\u003Clens-id>.md` — `specs\u002F\u003Cfeature>\u002Fworkshop\u002F\u003Clens-id>.md` during specify\u002Fintake, or\n   `specs\u002F\u003Cfeature>\u002Fiterations\u002F\u003CNNN>\u002Fworkshop\u002F\u003Clens-id>.md` during design analysis; **(b) ONLY AFTER that\n   nonempty file exists**, write the lens's complete `lens-applicability.json` entry from step 6 with\n   `moved_on: true`; **(c)** refresh the rolling handover through the core save path by running the handover\n   provider with `--source workshop` (one line; the SAME path the hooks use):\n   `pwsh -NoProfile -File .specify\u002Fextensions\u002Fspecrew-speckit\u002Fscripts\u002Fspecrew-handover-provider.ps1 --project-root . --source workshop`\n   — this captures the freshly-written `workshop\u002F` files into the handover so a resuming session inherits the\n   progress. (On the Claude host the `PostToolUse` hook ALSO refreshes the handover automatically the moment you\n   write the lens record; this explicit call is the cross-host fallback where PostToolUse is not wired.) A\n   session that resumes then reads the handover + the `workshop\u002F` folder and **continues from the next\n   un-persisted lens instead of restarting the workshop**. A lens that lives only in the chat scrollback is lost\n   on exit; a persisted lens is not. Then state which lens is next and reload this conduct + that lens's md.\n8. **Preflight the completed workshop before the spec writer or any boundary commit.** After the final lens is\n   durably complete, run exactly:\n   `pwsh -NoProfile -File .specify\u002Fextensions\u002Fspecrew-speckit\u002Fscripts\u002Fsync-boundary-state.ps1 -ProjectPath . -BoundaryType specify -PreflightOnly`.\n   This invokes the real specify gate without advancing lifecycle state. It validates binding tokens,\n   product-domain shape\u002Fprovenance, `implementation-rules.yml`, and runnable reviewer authorization. If it\n   refuses, repair the named workshop artifact, rerun the preflight, and re-ask the current question when the\n   durable decision changed. Only a successful preflight permits `speckit.specify`. Commit the boundary artifacts\n   only after the spec and checklist are written and the normal boundary sync succeeds.\n\n## When to Use\n\n- At `speckit.specify` \u002F intake, before the specify boundary syncs — run the lens workshop so the spec is\n  lens-informed.\n- At the design-analysis stop — co-design the architecture before `plan.md`.\n- **At the start of each new lens** — re-invoke for fresh, focused conduct.\n\n## Review Standard\n\nThis skill is only doing its job when, in the human's view of the conversation: each applicable lens was a\ngenuine discussion (not a one-shot MCQ); diagrams were SEEN in-band (ASCII inline or a clickable link); the\ncomponents were named with responsibilities and co-designed before options; and every agreement (lenses,\nco-design map, UI layout) was written to an artifact, not left in the scrollback.\n","# specrew-design-workshop\n\n**Type**: Facilitation Skill\n**Schema**: v1\n**Status**: Active design-lifecycle conduct\n\n## Purpose\n\nDrive the design lenses as a real, point-of-use **workshop** — not a checklist skimmed once at launch. Each\ntime you reach a lens, this skill is loaded fresh so the conduct is in focus exactly when you use it. **It is\nself-contained: one load gives you everything you need for the current lens.** Because most hosts do not\nguarantee a fresh reload per turn, **re-invoke this skill (or re-read it) at the start of each new lens.**\n\n## The Big Picture (read this first, every time)\n\nYou are facilitating a design conversation with a human, **one lens at a time**. The lenses live in\n`extensions\u002Fspecrew-speckit\u002Fknowledge\u002Fdesign-lenses\u002F` (deployed under the project's design-lens catalog):\n`architecture-core`, `component-design`, `requirements-nfr`, `ui-ux`, `data-storage`, `security-compliance`,\n`integration-api`, `devops-operations`, `observability-resilience`, `code-implementation`. For each lens you work:\n\n1. **Load that lens's md** (`design-lenses\u002F\u003Clens-id>.md`) for its `## Design Decision Points` and\n   `## Workshop Conduct` — that is the lens's focused agenda. Do not improvise it from memory.\n2. **Facilitate** the discussion for that lens (the method below).\n3. **Record** the decisions + agreement.\n4. **Re-invoke this skill** and go to the next lens.\n\nThe per-lens *knowledge* is in the lens md; this skill is the *method* that ties the lenses together. Keep\nboth in view.\n\n**Render before you ask — the question may only reference what is on screen (A6\u002FA7\u002FFR-037).** Before you\nraise ANY confirm\u002Fapprove question (as a typed \"does this work? \u002F move on? \u002F approve\" prompt) about the lens\n**agenda + depths**, a per-lens diagram, the\ncomponent map, an options\u002Ftrade-off set, or a design verdict — the thing you are asking about MUST already be\n**rendered in your assistant message in THIS exchange**, in prose \u002F console-ASCII the human can see. The\nquestion may reference ONLY content that is on screen. **Never** ask from a count (\"8 lenses\", \"13\ncomponents\"), a summary (\"the agenda\", \"the map\"), or a \"shown above \u002F as proposed\" that was not actually\nshown. The failure mode this exists to stop is **menu-before-render**: render + explain, THEN ask. **Be verbose**\n— explain the agenda \u002F diagram \u002F map \u002F options fully in prose first, and make the question and choices\nself-explanatory (not a terse \"does this work? — 8 lenses\").\n\n**Typed-turn authority rule — every host:** workshop questions are visible prose and the human answers by typing.\nOffer known choices as a **numbered prose list answered by typing** (or ask concise free-form prose); do not call a\nstructured question\u002Fmenu tool anywhere in the product-domain phase, agenda confirmation, or lens work. A picker\ncan swallow rendered context. **Ctrl+O is no answer**: Copilot reports it as `User skipped question`, which a model\nthen misread as delegation. Neither dismissal nor a picker selection is workshop authority. If one is invoked anyway,\nre-render the unanswered question as prose and wait: never infer defaults, `human-delegated`, or `human-skipped`.\nClaude additionally removes `AskUserQuestion` at capability level. The other hosts are held by the hook-owned\ntyped-turn receipt: no typed reply means no receipt and the controller\u002Fboundary refuses the claimed agreement.\nAfter each typed reply, read the latest matching line from `.specrew\u002Fruntime\u002Fworkshop-authority.jsonl` and copy\nits `receipt_id` into the record you are about to write. If there is no matching receipt, re-render the question\nand stop; never manufacture an id or edit the hook-owned store.\n\n**The `file:\u002F\u002F\u002F` links go in your prose before the question, too (dogfood finding).** When a confirm\u002Fapprove\nquestion references or asks the human to review an artifact — the spec, a lens workshop record, the\ndesign-analysis, a diagram file — emit the **bare clickable `file:\u002F\u002F\u002F` links in your assistant message in THIS\nexchange**, never only inside structured question fields and never as a bare \"see the file above\". Structured\nquestion UIs may not linkify `file:\u002F\u002F\u002F`; Claude's picker is disabled for this skill because it can drop the\nwhole preceding message. Rule 14A's clickable-`file:\u002F\u002F\u002F` guarantee must therefore hold *at* the confirmation,\ncarried by visible prose.\n\n**(A8\u002FFR-041) Open each lens with a presentation + an open question, never a menu** (The Method step 3) — this is\nwhat actually rendered the lens content on Claude in the dogfood (the per-lens open has no competing menu, so\nthe content cannot collapse into one). **Governing model (dogfood-proven):** content whose next move is *open\ndiscussion* renders reliably. On Claude, content immediately before `AskUserQuestion` can be swallowed by the\npicker; the deployed skill therefore removes that tool for the whole workshop. This is a capability-level\nguard, not another conduct instruction or hidden marker. An earlier catalog-at-open front-load was reverted\nbecause it still skimmed on Claude and was redundant on prose hosts that render the agenda inline.\n\n## First stage — the product-domain phase (run before everything, Feature 176)\n\nBefore the lens applicability agenda (The Method below), run the **product-domain** first-stage\nphase — the pre-technical product\u002Fproblem grounding. It is **always applicable** and runs FIRST,\nbefore any technical lens or the applicability questionnaire. It is NOT a row in\n`applicability-map.json`; it is the first-stage phase ahead of the deterministic selector.\n\n0. **Establish durable pre-agenda controller state before the first product-domain question.** Immediately\n   after the feature scaffold creates its nonempty `spec.md`, and only when the feature-level\n   `lens-applicability.json` is absent, run:\n   `pwsh -NoProfile -File .specify\u002Fextensions\u002Fspecrew-speckit\u002Fscripts\u002Finitialize-workshop-controller-state.ps1 -ProjectRoot . -FeatureRef \u003Cexact-feature-ref>`.\n   The helper atomically writes the strict feature-only `agenda_status: pending-confirmation` shape with empty\n   `selected` and `workshop` collections. It refuses to overwrite an existing artifact. On resume, read and\n   honor the existing exact-scoped artifact instead of rerunning the helper. This is controller truth for the\n   product-domain and agenda-confirmation questions; it is not a hidden model marker, environment variable, or\n   transcript heuristic.\n1. **Load the lens** `design-lenses\u002Fproduct-domain.md` for its decision areas, depth model,\n   evidence vocabulary, run cadence, and conduct.\n2. **Select the depth** (Light \u002F Standard \u002F Deep) by **risk and novelty**, and say why. A tiny\n   utility gets a Light pass that records why deeper discovery is not warranted.\n3. **Reframe a solution-first request into the problem.** When the human asks to \"build X\", surface\n   who it is for, the pain\u002Fjob, the MVP, the non-goals, and the constraints BEFORE any design. If the\n   requested feature is not aligned to the pain \u002F MVP \u002F constraints \u002F alternatives, surface that\n   before plan.\n4. **Capture the product context** at the selected depth (users and stakeholders; pain\u002Fjob and\n   current workaround; existing system\u002Fcontext; constraints; outcomes; MVP\u002Fnon-goals\u002Fvision;\n   alternatives at Standard\u002FDeep), and **tag every material statement** with its evidence quality\n   (`known` \u002F `assumed` \u002F `unknown` \u002F `research-needed`; a `research-needed` carries `load_bearing`).\n   Honest tags prevent confident product fiction.\n5. **No batch confirmation (FR-009).** The phase cannot be satisfied by approving an agenda or a batch\n   \"confirm all\". It needs scoped product-domain confirmation, or an explicit, honestly recorded\n   \"you decide\" \u002F \"skip\". Record the provenance (`human-confirmed` \u002F `human-delegated` \u002F\n   `human-skipped`) and its matching scope.\n6. **Persist both records** under `specs\u002F\u003Cfeature>\u002Fworkshop\u002F`: `product-domain.yml` (structured,\n   schema-validated against `contracts\u002Fproduct-domain.schema.json`) and `product-domain.md`\n   (human-readable), and summarize the decisions into `spec.md` — not as the sole source. The\n   specify-gate floor REQUIRES this for a substantive feature before specify syncs; an absent record\n   is surfaced, never silently skipped.\n\n**Run cadence**: the product-domain phase runs before EVERY feature at adaptive depth — not once. In\nV1 every feature is `context_scope: feature_standalone`; once Proposal 162 ships, later features run\nin delta mode (`feature_delta`) against the inherited product baseline (`product_baseline`).\n\nOnly after the product-domain phase is captured do you move to the lens applicability agenda below.\n\n## The code-implementation lens (auto-on for code features; Feature 177)\n\nThe `code-implementation` lens captures *how the code is written* (implementation craft) as binding\nconstraints for implement. It is **always-applicable for any feature that writes code** (auto-on; skip\nonly doc-only \u002F config-only slices, with a recorded reason). It is **conduct-driven** — NOT a row in\n`applicability-map.json` (drift D-001), exactly like `product-domain`: you include it for code features\nwithout a yes\u002Fno applicability question. Work it **after the technical lenses** (it depends on the\nresolved stack + the architecture decision). Load its md\n`design-lenses\u002Fcode-implementation.md` for the decision spine, per-stack dilemmas, the grouping model, the\nrun-cadence, and the full conduct. The conduct in brief:\n\n1. **Source of code-rules truth FIRST** (the Figma-equivalent question): ask whether the human has an\n   existing coding guideline OR one or more **example projects** to emulate (a GitHub repo, a local path,\n   or other) for code style, language constructs, and patterns — or none.\n2. **Assisted ingestion** (when a guideline \u002F example project is provided; agent-reasoning, no parser):\n   map it onto the `code-rules.yml` catalog (auto-check matches, flag conflicts for the human), and extract\n   non-catalog conventions as custom rules with provenance (`from-guideline` \u002F `from-example-project` +\n   the source ref). Company\u002Forg-level rules persist to the reusable project overlay `code-rules.local.yml`\n   (additive + per-rule override; never drops a shipped rule).\n3. **Resolve the stack**, then present the **grouped, pre-checked set\u002Funset checklist** — baseline stated\n   as a summary (exceptions only), the consequential decision-prompts paced (offer all-at-once OR\n   one-at-a-time), applicability-filtered rules shown only when their context applies. **Never a flat wall.**\n   **\"Use the defaults\" means one coherent bundle, not every familiar mechanism at once.** Check that the\n   chosen defaults can coexist before recording them. For example, a small manually composed .NET CLI may own\n   one shared `HttpClient`; choosing `IHttpClientFactory` implies factory\u002Fservice-provider composition and must\n   not be paired with \"manual `new()`, no DI container\". Resolve contradictions with the human in this lens.\n4. **Dependency selection (FR-013)**: present **\"use existing project tools \u002F no new dependency\" first**\n   plus options; for any chosen dependency capture version, license, source org, canonical URL, maintenance\n   signal, security\u002Fadvisory status, compatibility, cost\u002Fquota, coupling weight, replaceability, and test\n   implications into the manifest `dependency_policy`.\n5. **Capture** the selections\u002Fdecisions\u002Fcustom-rules\u002Fdependency-policy by authoring the per-feature\n   `implementation-rules.yml` manifest (reference-by-ID), schema-validated against the manifest schema\n   `implementation-rules.schema.json` (beside the lens in the design-lens catalog), plus the human-readable\n   `workshop\u002Fcode-implementation.md`; record the lens in `lens-applicability.json`. (Like `product-domain`,\n   you author the record by hand following the schema -- you do NOT call a PowerShell writer; the lens\n   helper validates, it is not invoked mid-workshop.) Reviewer setup is part of this capture: list the live\n   hosts, present the marked default, record one concrete host + authorization reference, and run the explicit\n   `specrew review --host ... --authorization-ref ...` setup command. Never leave a completed code lens with\n   an unresolved `auto-select`; that state cannot launch a reviewer in a fresh project.\n6. **Run cadence**: the rules are mostly product-level — decide once at a product-level workshop, inherit\n   per feature, re-open only the parts a new technology or programming language changes\n   (`context_scope` hooks; V1 `feature_standalone`; forward-compatible with Proposal 162).\n\nThe **implement-time half** is the separate `specrew-code-rules` skill, which reads this manifest while the\ncoding agent writes code and surfaces the rules task-scoped. The acceptance gate is the deployed dogfood\n(SC-004\u002FSC-007\u002FSC-008), not unit-green.\n\n## The Method (the same for every lens)\n\n1. **Frame the phases + hand over the agenda (A6\u002FFR-034, A7\u002FFR-040).** Tell the human up front: the workshop\n   *gathers* inputs and constraints; the system **structure** (components, responsibilities, flows) is designed\n   *with* them at the design-analysis step — not decided in intake. So they know the collaboration is coming.\n   **Before the heavy prep, keep them oriented (FR-040):** say plainly that you are *preparing the workshop and\n   it will take a moment* (so a pause does not read as a hang), and hand them the **agenda as an assignment** —\n   list the lenses you will work and, for each, the decision it will ask of them — so they can think or research\n   while you load. The wait becomes preparation, and a prepared human engages per-lens (which is what keeps the\n   integrity rule in step 6 honest). **Tell the human they can just talk (dogfood finding):** in the same\n   framing, say plainly that at any lens, if a question is unclear, they want to open a file, or they need more\n   detail, they can simply *type* it (for example \"explain more\") instead of picking a menu option — you will\n   explain, then re-ask. Typed\n   conversation is the required path on every host; do not use the host's structured picker during the workshop.\n   **Product-domain owns product facts, not technical decisions.** The Crew proposes the product-domain depth\n   from risk and novelty and the human confirms or adjusts that depth in their typed product-domain reply; do not persist\n   the selected product-domain depth before that typed human confirmation. Product-domain may ask whether a\n   technology is already externally fixed as a constraint. It must not ask the human to choose runtime\u002Fstack,\n   output format, timeout\u002Freachability semantics, concurrency, or dependencies: those belong to their selected\n   technical lens, where each receives its own turn and typed receipt.\n2. **Infer applicability, then confirm (A4\u002FFR-025).** Propose which lenses apply WITH your reasoning; ask the\n   human only to confirm or adjust. Never make them answer obvious yes\u002Fno applicability; never silently\n   auto-resolve a material area. **Render the agenda IN-BAND before asking for confirmation — fill this\n   template; do NOT cram the lens list into a question UI** (a prose \"render first\" gets skimmed on some hosts;\n   the filled template is what the human reads while you prepare):\n\n   ```text\n   Workshop agenda — \u003CN> lenses\n\n   \u003Clens-id> (\u003Cfull | medium | light>) — \u003Cthe decision this lens will ask you to make for THIS feature>\n   \u003Clens-id> (\u003Cfull | medium | light>) — \u003Cthe decision …>\n   ...\n   Skipped: \u003Clens-id> — \u003Cwhy it does not apply here>\n   ```\n\n   Fill ONE line per applicable lens with its depth and the **concrete decision it raises** (not just the lens\n   name); render the whole filled block in your message, THEN ask the human to confirm or adjust it as visible\n   prose. Wait for a typed response on every host.\n   **Show the negative selection too.** One `Skipped:` line is required for **every** technical lens that is not\n   selected, with a feature-specific reason. Then ask, in plain language, whether the human wants to confirm or\n   change the selection. The first lens must not open on the same turn as the agenda: typed human confirmation is\n   the transition. A count such as \"5 lenses\" is not reviewable unless the omitted set is visible too.\n   **Render the authority-bound agenda before asking.** Build one JSON object shaped as\n   `{\"selected\":[{\"lens\":\"architecture-core\",\"depth\":\"medium\",\"decision\":\"...\"}],\"skipped\":[{\"lens\":\"ui-ux\",\"reason\":\"...\"}]}`,\n   then invoke `& .specify\u002Fextensions\u002Fspecrew-speckit\u002Fscripts\u002Fconfirm-workshop-agenda.ps1 -ProjectRoot . -FeatureRef \u003Cfeature> -AgendaJson $agendaJson -RenderOnly`.\n   Send the command's complete output as one unchanged, contiguous block. A short introduction or closing is\n   allowed, but may not replace, paraphrase, or split the canonical block. The hook binds the typed reply to the\n   structured agenda digest after proving that complete block was visible. Wait for the human's\n   typed response. **The moment they confirm, PERSIST it before opening lens 1** by invoking the same\n   command with the identical `$agendaJson` and without `-RenderOnly`. If the selected set, depths, decisions,\n   skipped set, or reasons change, render the new canonical agenda and ask again; an older receipt cannot authorize it.\n   The writer changes `agenda_status` to `confirmed` only when selected + skipped cover the whole deployed\n   technical-lens catalog and records `agenda_confirmation: human-confirmed` with\n   `agenda_confirmation_scope: lens-selection`. It also requires hook-owned typed reply receipts for product-domain\n   and the complete agenda; Ctrl+O or a picker result produces neither. It refuses incomplete coverage or missing\n   authority. The per-lens `workshop` records\n   are added later, as each lens completes per step 6. A resume can only compute the remaining agenda if the agenda\n   itself is on disk; an agenda that lives only in scrollback is lost on exit (observed: an unpersisted agenda made\n   a resuming host re-run specify instead of continuing the workshop).\n   If the named action does not clear a refusal on the next attempt, stop; do not retry again and do not write\n   workshop records by hand to get past it. Report calmly in the user's project terms: what you were doing; that\n   their answers are safe and nothing has been lost (only when that is true); what you could not complete without\n   diagnosing or assigning fault; one concrete next action you can perform; and ask the human for their approval.\n   Keep machinery diagnosis in the drift record, not in the message the human must act on.\n   When the human approves preparing a repair for inconsistent pre-agenda progress, use the governed recovery\n   surface instead of editing `lens-applicability.json`: run\n   `pwsh -NoProfile -File .specify\u002Fextensions\u002Fspecrew-speckit\u002Fscripts\u002Frepair-workshop-controller-state.ps1 -ProjectRoot . -FeatureRef \u003Cexact-feature-ref> -Request`,\n   render its complete proposal, and wait for the exact typed authorization it names. Only after that reply has\n   been captured may you run the same command with `-Apply`. The repair is state-hash-bound, preserves the saved\n   product-domain records, writes an audit record, and refuses if anything changed after authorization.\n   **Agenda confirmation is not lens-question confirmation.** This confirm point approves only the selected\n   lens list and depths. It does NOT answer the lenses. Do NOT offer or accept a batch shortcut such as \"Confirm\n   all as proposed\", \"approve all lens decisions\", or \"use the proposed decisions for every lens\" as\n   `human-confirmed` \u002F `lens-question`. After the agenda is confirmed, every selected lens still needs its own\n   lens-specific turn and its own human answer, explicit delegation, or explicit skip.\n3. **Per-lens facilitated discussion — open with a presentation + an OPEN question, never a menu first (A4\u002FFR-025, A8\u002FFR-041b).**\n   The **first turn of every lens** MUST be you *presenting* the lens — its decision points, and as it develops\n   its diagram \u002F component map — followed by an **open, free-text question** (\"how should we approach this?\",\n   \"what are your constraints?\"). Do **NOT** open a lens with an `AskUserQuestion` \u002F structured menu: that is the\n   move that lets the content collapse into the menu's question field and never get rendered (the A8\n   `AskUserQuestion` tool-gravity failure). On Claude the tool is unavailable for the entire skill; use visible\n   prose and typed numbered choices. On every other host the same choices remain visible prose with a typed reply;\n   do not switch to a structured menu after rendering. Binary\n   test: did this lens open with a rendered presentation, or with a menu? Open with the presentation.\n   **One selected lens = one lens turn.** Do NOT bundle several selected lenses into one combined presentation\n   and one \"confirm all\" question, even for a tiny feature or light-depth lenses. A lens turn may summarize the\n   already-confirmed agenda, but it must focus on exactly one lens's decision points, ask for that lens's answer,\n   and wait for the human (or an explicit \"you decide for this lens\" \u002F \"skip this lens\") before moving on.\n   This is also a question-content boundary: a product-domain turn must not harvest runtime\u002Fstack, output,\n   timeout\u002Freachability, concurrency, or dependency decisions owned by later technical lenses. A receipt for\n   one lens cannot authorize another lens's answer.\n   **Keep the controller-owned workshop state durable and complete (FR-055\u002FFR-056).** Before you stop and wait\n   for an answer, ensure the applicability artifact for the CURRENT workshop scope exists and reflects either\n   the strict feature-level pre-agenda state or the current confirmed agenda. During specify\u002Fintake, authority is\n   the feature-level\n   `specs\u002F\u003Cfeature>\u002Flens-applicability.json`; during design analysis it is the exact iteration's\n   `specs\u002F\u003Cfeature>\u002Fiterations\u002F\u003CNNN>\u002Flens-applicability.json`. Never invent an iteration during feature intake.\n   Do not rely on a model-authored hidden marker, an environment variable, or a host question-tool transcript:\n   hosts can omit, transform, or swallow those surfaces. The Stop controller derives `active` only from the exact\n   scoped artifact: either its strict feature-only `pending-confirmation` shape with a valid scaffolded spec, or\n   its confirmed nonempty, unique selected agenda and strict ordered completion records. While that state is\n   valid and incomplete, ordinary product-domain, agenda, and lens questions remain conversational and the\n   generic five-section non-boundary packet is suppressed; a real lifecycle boundary still has precedence.\n   **Finish each lens durably in this order:** write the nonempty `workshop\u002F\u003Clens-id>.md` decision record first,\n   then persist that lens's full step-6 entry with `moved_on: true`. On the final selected lens, that second write\n   makes the workshop `complete` and ordinary Stop behavior resumes immediately. A loose flag, missing record,\n   duplicate\u002Fout-of-order lens, malformed artifact, or mismatched confirmation is invalid and cannot keep the\n   exception active. Never render the generic five-section packet merely to ask or answer an ordinary lens\n   question; use the six-section packet only for a genuine lifecycle boundary.\n   **Carry binding decisions across lenses.** Before opening a new lens, reread the completed records' durable\n   `bindings`. A delegated\u002Fdefault decision may fill an unresolved gap, but it MUST NOT contradict an earlier\n   human-confirmed binding. If the human intentionally changes a binding, reconcile every affected lens record\n   to the same value before moving on; never leave two durable answers for one architectural fact.\n   **Pace a dense lens — after presenting, you MUST offer all-at-once OR one-at-a-time (A8\u002FFR-041b UX, every host).**\n   A lens with several decision points (architecture-core, component-design, security-compliance) lands as an\n   overwhelming **wall** if you present everything and end with one open question that secretly bundles five\n   subjects — and a numbered list of subjects reads like pickable options when it is not. So on a dense lens\n   (**three or more** decision points) your closing move **MUST** offer the human a **pacing choice** — *answer it\n   all in one go, or have me take you through the decisions one at a time* (then ask decision 1, wait for the\n   answer, then decision 2, …, each as its own focused question). This is **not optional on a dense lens and\n   applies on every host** — a single open question covering five subjects is as hard on Copilot\u002FCodex\u002FAntigravity\n   as on Claude; do not bundle. Respect their pick; never force the whole wall on someone who wants to step\n   through it. A light single-decision lens skips the offer — just ask the one open question. Then, for\n   the current lens, raise its decision points (from its\n   md), offer options where useful, capture the human's needs + decisions + explicit agreement, and **iterate\n   until the human says \"move on\"** before the next lens. Adapt depth to the user-profile expertise dials\n   (concise where high; explain + recommend a default where low). Right-size — not a fixed nine-lens marathon.\n   **Match the question FORM to the question**: for a discrete, enumerable choice (e.g. decomposition\n   vocabulary — IDesign \u002F Clean Architecture \u002F modular; one service vs split; fixed vs open taxonomy), spell\n   out every option and an explicit \"other \u002F let me explain\" path so the human can pick fast. This is a numbered\n   prose list answered by typing on every host. For a genuinely open question, discuss in prose. Both are fine — do not\n   force a discrete pick into long prose, nor an open design question into a rigid one-shot MCQ. **Surface EVERY\n   selected lens to the human and get a real answer before you record it (A7\u002FFR-038):** intake is NOT \"specific\n   enough\" until each selected lens has either the human's confirmation OR an explicit \"you decide \u002F skip\" from\n   them. You may NOT decide after a few questions that intake is done and then write up the remaining lenses\n   yourself — that is the exact failure this rule exists to stop. When you move to the next lens (loading it\n   lazily), announce it so the pause is legible: *\"preparing lens X of N: &lt;lens&gt; — get ready, this one\n   decides …\"* (FR-040).\n4. **Surface visuals IN-BAND so the human can SEE them (A5\u002FA6\u002FFR-030–FR-031\u002FFR-037).** On a terminal\u002Fconsole\n   host a fenced ```mermaid``` block is **source text, not a rendered picture** — only **console ASCII**\n   actually renders inline. So the diagram you show the human MUST be **console ASCII art rendered directly in\n   your message** (boxes, arrows, labels) — reach for ASCII first and by default. Use mermaid \u002F svg \u002F html only\n   as a *richer, additional* artifact written to a file under the workshop folder — and when you do, you MUST\n   emit its clickable `file:\u002F\u002F\u002F` link in the SAME message. A diagram the human cannot see is NOT surfaced;\n   never let the human reach a decision while a diagram lives only in a file or only as mermaid source. A\n   per-lens diagram is **expected** for the structural lenses (architecture, data, security, integration,\n   devops) and any UI-bearing feature (ui-ux: a layout sketch). Intake is bidirectional: offer to plot the\n   lens's diagram AND ask whether the human has an existing diagram\u002FFigma\u002Fwhiteboard photo to bring. **At ANY\n   approval point — \"approve\" \u002F \"move on\" \u002F \"co-design\" \u002F \"does this work as the baseline\" — the diagram or\n   component map MUST be rendered in console-ASCII IN THE SAME MESSAGE as the question; NEVER stand in a\n   reference to it (a file path, or a bare count like \"13 components\" \u002F \"4 managers\"). The human approves what\n   is on screen, not what is in a file.** Persist the keeper to the workshop file AFTER and IN ADDITION TO the\n   in-band render (step 6) — never INSTEAD of it. (Empirically: testLenses8 on the Claude host wrote the\n   component map to the workshop file and asked the human to approve \"13 named components\" it never showed,\n   while Copilot and Antigravity rendered it in-band first — writing the file is not surfacing.)\n5. **Co-design — do NOT hand down finished options (A6\u002FFR-035\u002FFR-036).** At the design-analysis stop:\n   - **Co-decide the design method \u002F decomposition style** (DDD bounded-contexts, IDesign volatility-based,\n     modular monolith, microservices, layered) as an expertise-adapted discussion; record the choice as a\n     binding constraint. Not a bare multiple-choice prompt; do not silently assume it.\n   - **Co-build the component map — render the FULL form, never a summary or a count.** Whenever you present\n     the component map (at the component-design lens AND at the design-analysis stop), put this IN YOUR MESSAGE,\n     in order: **(1)** a **console-ASCII diagram** of all components on their layers with dependency arrows;\n     **(2) then a named list grouped by the decomposition vocabulary** you bound at the design-method step —\n     Managers \u002F Engines \u002F ResourceAccessors for IDesign; bounded contexts \u002F aggregates \u002F entities for DDD;\n     layers for layered; services for microservices — with **EVERY component named and its one-line\n     responsibility** (e.g. `TrayClient — owns the tray UI + global hotkey`). **NEVER a bare count** (\"6\n     resource accessors\", \"3 engines\") and never a \"map above\" \u002F file reference — list every component by name.\n     ONLY after the diagram + the full named list are on screen do you ask the human to approve \u002F rename \u002F\n     split \u002F merge \u002F reassign, and walk at least one key **flow** (user + system actions) through it together.\n     **If they ask for a change, re-render the updated diagram + list, then ask again** — iterate until the\n     human agrees the decomposition, responsibilities, and flows are right. **Use this fill-in template so the\n     form is consistent** — copy it into your message and replace every `\u003C...>`, list EVERY component (never a\n     count), group by your chosen vocabulary, render it in-band BEFORE the approve\u002Fchange question, and\n     re-render the filled template on any change:\n\n     ```text\n     Proposed component map\n\n     \u003Cconsole-ASCII diagram: every component on its layer, dependency arrows pointing inward>\n\n     \u003Cvocabulary group 1 — Managers | Bounded Contexts | Layers | Services>:\n       \u003CComponentName> — \u003Cone-line responsibility>\n       \u003CComponentName> — \u003Cone-line responsibility>\n     \u003Cvocabulary group 2 — Engines | Aggregates | ...>:\n       \u003CComponentName> — \u003Cone-line responsibility>\n     \u003Cvocabulary group 3 — Resource Accessors | Repositories | ...>:\n       \u003CComponentName> — \u003Cone-line responsibility>\n\n     Key flow: \u003Cactor action> -> \u003CComponent> -> \u003CComponent> -> \u003Coutcome>\n     ```\n\n   - **Only then present the remaining trade-off options** (transport, store technology, granularity) for the\n     human's verdict — the leftover consequential choices, not the whole design handed down.\n   - Where the human's architecture expertise is low you MAY drive the decomposition and explain more, but you\n     MUST still confirm the responsibilities and flows with them, not author them silently.\n6. **Capture the agreements (A4\u002FA6\u002FSC-021\u002FSC-025).**\n   - Per-lens workshop record in the feature-level `lens-applicability.json` — set `workshop_intake: true`,\n     `confirmation_required: true`, `agenda_status: confirmed`, the `selected` lens-id list, and a SINGLE\n     top-level `workshop` object\n     **keyed by lens id**, each value carrying the EXACT fields the gate checks: `agenda` (array of questions\n     raised), `decision` (a SINGLE STRING summarizing the decision + agreement), `depth`, `moved_on: true`, and\n     **`confirmation`** — the provenance, one of `human-confirmed | human-delegated | human-skipped` (A7\u002FFR-039,\n     SC-026), plus **`confirmation_scope`** — `lens-question` for `human-confirmed`, `explicit-delegation` for\n     `human-delegated`, or `explicit-skip` for `human-skipped`; and **`human_turn_receipt`**, the receipt id from\n     the latest matching line in `.specrew\u002Fruntime\u002Fworkshop-authority.jsonl`. No receipt means re-render and wait.\n     Exact shape — get it right the first time:\n\n     Record every load-bearing or cross-lens decision in an optional `bindings` object using stable lowercase\n     keys and token values. Reuse the same key whenever another lens touches that decision. Repeated values must\n     match exactly; the controller stops on a conflict before the workshop can finish.\n\n      ```json\n      { \"workshop_intake\": true, \"confirmation_required\": true, \"agenda_status\": \"confirmed\", \"selected\": [\"architecture-core\"],\n        \"workshop\": { \"architecture-core\": { \"agenda\": [\"q1\",\"q2\"], \"decision\": \"what was decided + agreed\", \"depth\": \"full\", \"moved_on\": true, \"confirmation\": \"human-confirmed\", \"confirmation_scope\": \"lens-question\", \"human_turn_receipt\": \"\u003Cmatching receipt_id>\", \"bindings\": { \"article-initiation\": \"on-demand\" } } } }\n      ```\n\n     It is `workshop` -> `\u003Clens-id>` -> fields (NOT `\u003Clens-id>` -> `workshop`), and `decision` is a singular\n     string, NOT a `decisions` array — the inverted nesting or a `decisions` array FAILS the SC-021 gate. Extra\n     fields (such as the `diagram` reference below) are fine. **Persist each keeper diagram you render for a lens** (the trust-boundary diagram, the ERD, the\n     component map, the sequence, etc.) — write the ASCII (or mermaid) to a workshop folder,\n     `specs\u002F\u003Cfeature>\u002Fworkshop\u002F\u003Clens-id>.md`, and set that lens's `diagram` field to a **reference to that\n     file** (the path + a one-line caption), NOT a prose description. A diagram that lives only in the chat\n     scrollback is lost; the workshop folder makes the design reviewable from the artifacts.\n   - **Integrity — never manufacture agreement (A7\u002FFR-038, SC-026).** Set `confirmation: human-confirmed` and\n     `confirmation_scope: lens-question` ONLY for a lens whose substantive workshop questions you actually\n     surfaced and the human confirmed. Lens approval is not workshop-question approval. If the human explicitly\n     said \"you decide\" or \"skip\" for that lens's questions, set `human-delegated` + `explicit-delegation` or\n     `human-skipped` + `explicit-skip` and say so honestly in `decision` — do NOT write a fabricated \"the human\n     agreed to X\" for a lens they never saw. **A batch confirmation is not valid provenance:** if the human says\n     \"confirm all as proposed\" at the agenda or summary level, that confirms only the agenda\u002Fscope and cannot\n     produce `human-confirmed` \u002F `lens-question` for any lens. If the human explicitly delegates a shortcut such\n     as \"you decide the remaining lenses\", record each affected lens as `human-delegated` +\n     `explicit-delegation`, not `human-confirmed`. **Count self-check before you record:** you are about to write\n     N lens records — you must have asked, or been explicitly told to decide\u002Fskip, N times, one lens at a time.\n     A dismissed picker is an absence, never explicit delegation. If you stopped early, go back and surface the rest; you may not declare intake \"specific enough\" and fill in\n     the remaining lenses yourself. The SC-026 gate blocks the specify boundary until every selected lens carries\n     a `confirmation` and matching `confirmation_scope` — but it cannot see transcript truthfulness; that\n     integrity is on you, and the Squad re-dogfood checks it.\n   - At design-analysis, a `## Co-Design Record` in `design-analysis.md` with the agreed\n     component-to-responsibility map + at least one agreed flow + a human-agreed marker, and — when ui-ux is in\n     scope — the **agreed UI\u002Fscreen layout** (the ASCII sketch the human approved). Set `co_design: true` in the\n     iteration's `lens-applicability.json` so the deterministic co-design-record floor applies. An agreement\n     that lives only in the chat scrollback is lost.\n7. **Checkpoint this lens durable, THEN re-invoke for the next so a mid-workshop exit\u002Fswitch is recoverable.**\n   The workshop is long, and an exit or host-switch mid-workshop is expected, not exceptional — so make each\n   lens durable the moment you finish it, never \"all at the end\". BEFORE you move to the next lens:\n   **(a) persist the Markdown decision record FIRST** at the exact current scope's\n   `workshop\u002F\u003Clens-id>.md` — `specs\u002F\u003Cfeature>\u002Fworkshop\u002F\u003Clens-id>.md` during specify\u002Fintake, or\n   `specs\u002F\u003Cfeature>\u002Fiterations\u002F\u003CNNN>\u002Fworkshop\u002F\u003Clens-id>.md` during design analysis; **(b) ONLY AFTER that\n   nonempty file exists**, write the lens's complete `lens-applicability.json` entry from step 6 with\n   `moved_on: true`; **(c)** refresh the rolling handover through the core save path by running the handover\n   provider with `--source workshop` (one line; the SAME path the hooks use):\n   `pwsh -NoProfile -File .specify\u002Fextensions\u002Fspecrew-speckit\u002Fscripts\u002Fspecrew-handover-provider.ps1 --project-root . --source workshop`\n   — this captures the freshly-written `workshop\u002F` files into the handover so a resuming session inherits the\n   progress. (On the Claude host the `PostToolUse` hook ALSO refreshes the handover automatically the moment you\n   write the lens record; this explicit call is the cross-host fallback where PostToolUse is not wired.) A\n   session that resumes then reads the handover + the `workshop\u002F` folder and **continues from the next\n   un-persisted lens instead of restarting the workshop**. A lens that lives only in the chat scrollback is lost\n   on exit; a persisted lens is not. Then state which lens is next and reload this conduct + that lens's md.\n8. **Preflight the completed workshop before the spec writer or any boundary commit.** After the final lens is\n   durably complete, run exactly:\n   `pwsh -NoProfile -File .specify\u002Fextensions\u002Fspecrew-speckit\u002Fscripts\u002Fsync-boundary-state.ps1 -ProjectPath . -BoundaryType specify -PreflightOnly`.\n   This invokes the real specify gate without advancing lifecycle state. It validates binding tokens,\n   product-domain shape\u002Fprovenance, `implementation-rules.yml`, and runnable reviewer authorization. If it\n   refuses, repair the named workshop artifact, rerun the preflight, and re-ask the current question when the\n   durable decision changed. Only a successful preflight permits `speckit.specify`. Commit the boundary artifacts\n   only after the spec and checklist are written and the normal boundary sync succeeds.\n\n## When to Use\n\n- At `speckit.specify` \u002F intake, before the specify boundary syncs — run the lens workshop so the spec is\n  lens-informed.\n- At the design-analysis stop — co-design the architecture before `plan.md`.\n- **At the start of each new lens** — re-invoke for fresh, focused conduct.\n\n## Review Standard\n\nThis skill is only doing its job when, in the human's view of the conversation: each applicable lens was a\ngenuine discussion (not a one-shot MCQ); diagrams were SEEN in-band (ASCII inline or a clickable link); the\ncomponents were named with responsibilities and co-designed before options; and every agreement (lenses,\nco-design map, UI layout) was written to an artifact, not left in the scrollback.\n",{"name":5,"description":7,"disallowed-tools":24},"AskUserQuestion","main","176da577bfd98859f793bdd1312690c6c3940c67",".claude\u002Fskills\u002Fspecrew-design-workshop","https:\u002F\u002Fgithub.com\u002Falonf\u002FKeyContextAI","https:\u002F\u002Fgithub.com\u002Falonf","http:\u002F\u002Falonfliess.me","CTO & Founder of Zionet.\r\nMicrosoft Regional Director and Foundry MVP.\r\n","2026-08-28 05:21:35",{"source":8,"skillName":5,"commands":34,"prompt":44,"zipAvailable":45},[35,38,41],{"tool":36,"command":37},"npx","npx skills add alonf\u002FKeyContextAI --skill specrew-design-workshop",{"tool":39,"command":40},"bunx","bunx skills add alonf\u002FKeyContextAI --skill specrew-design-workshop",{"tool":42,"command":43},"pnpm","pnpm dlx skills add alonf\u002FKeyContextAI --skill specrew-design-workshop","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Falonf\u002FKeyContextAI\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fspecrew-design-workshop\u002FSKILL.md",true,[47],{"path":48,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":51,"sourceHtmlUrl":52},".specrew-managed","OTHER","其他",80,"https:\u002F\u002Fgithub.com\u002Falonf\u002FKeyContextAI\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fspecrew-design-workshop\u002F.specrew-managed",[54,61,68,75,82,89],{"skillNo":55,"name":56,"nameZh":6,"description":57,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":58,"categories":59,"keywords":60,"publishedAt":19},"01M7AVHRWWXA0BQ5PXNA8FG30W","windows-compatibility","Cross-platform path handling and command patterns","https:\u002F\u002Fgithub.com\u002Falonf\u002FKeyContextAI\u002Fblob\u002Fmain\u002F.squad\u002Ftemplates\u002Fskills\u002Fwindows-compatibility\u002FSKILL.md",[],[],{"skillNo":62,"name":63,"nameZh":6,"description":64,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":65,"categories":66,"keywords":67,"publishedAt":19},"01M7AVHABPRYXYM3GWJY1RW92A","versioning-policy","Semver versioning rules for Squad SDK and CLI — prevents prerelease version incidents","https:\u002F\u002Fgithub.com\u002Falonf\u002FKeyContextAI\u002Fblob\u002Fmain\u002F.squad\u002Ftemplates\u002Fskills\u002Fversioning-policy\u002FSKILL.md",[],[],{"skillNo":69,"name":70,"nameZh":6,"description":71,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":72,"categories":73,"keywords":74,"publishedAt":19},"01M7AVH4KNP3TCVQDRVS70VP19","tiered-memory","Three-tier agent memory model (hot\u002Fcold\u002Fwiki) for context reduction per spawn","https:\u002F\u002Fgithub.com\u002Falonf\u002FKeyContextAI\u002Fblob\u002Fmain\u002F.squad\u002Ftemplates\u002Fskills\u002Ftiered-memory\u002FSKILL.md",[],[],{"skillNo":76,"name":77,"nameZh":6,"description":78,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":79,"categories":80,"keywords":81,"publishedAt":19},"01M7AVH18JDSA55Z7KK2GH6X49","test-discipline","Update tests when changing APIs — no exceptions","https:\u002F\u002Fgithub.com\u002Falonf\u002FKeyContextAI\u002Fblob\u002Fmain\u002F.squad\u002Ftemplates\u002Fskills\u002Ftest-discipline\u002FSKILL.md",[],[],{"skillNo":83,"name":84,"nameZh":6,"description":85,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":86,"categories":87,"keywords":88,"publishedAt":19},"01M7AVGVGWYD7NBNXAV4PAZJ92","squad","Squad's command catalog and interactive menu. Invoke via \u002Fsquad (slash command) or natural language (\"squad commands\", \"what can squad do\", \"show me squad options\"). Presents categorized operations (Install & Upgrade, Team Management, Issues & PRs, Plugins & Skills, Model & Cost, Sessions & State) as an interactive picker. Routes to the right squad CLI command or the Squad coordinator agent.","https:\u002F\u002Fgithub.com\u002Falonf\u002FKeyContextAI\u002Fblob\u002Fmain\u002F.squad\u002Ftemplates\u002Fskills\u002Fsquad\u002FSKILL.md",[],[],{"skillNo":90,"name":91,"nameZh":6,"description":92,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":93,"categories":94,"keywords":95,"publishedAt":19},"01M7AVGPSFT1WHW3PK35DHS5XK","squad-version-check","Internals of how @bradygaster\u002Fsquad-cli stamps its version, how `squad upgrade` works (what it preserves vs overwrites), and how to probe the npm registry for the latest version from a coordinator prompt.","https:\u002F\u002Fgithub.com\u002Falonf\u002FKeyContextAI\u002Fblob\u002Fmain\u002F.squad\u002Ftemplates\u002Fskills\u002Fsquad-version-check\u002FSKILL.md",[],[]]