[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M783DHYDXFS9PJN03SW72W8A":3,"skill-same:01M783DHYDXFS9PJN03SW72W8A":326},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":15,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":24,"commitSha":25,"skillDirectory":26,"forksCount":13,"repositoryHtmlUrl":27,"repositoryHomepageUrl":6,"ownerHtmlUrl":28,"ownerBlogUrl":6,"ownerDescription":29,"updatedAt":30,"install":31,"resources":44},"01M783DHYDXFS9PJN03SW72W8A","loop-builder",null,"Design and scaffold an agent \"loop\" — an unattended, scheduled, self-verifying agent workflow. Use this whenever the user wants to automate a recurring task, schedule an agent, run an agent unattended or overnight, set up monitoring, triage, or alerting, poll something on a cadence, or turn a manual repeated workflow into a self-running one — even if they never say the word \"loop.\" If a request implies \"do this every day \u002F on a schedule \u002F until some condition holds, without me typing each time,\" reach for this skill. It walks the seven-question blueprint, picks the simplest loop pattern, and scaffolds the six building blocks (schedule, isolation, skill, connectors, verifier, state) with a human-gate list and a budget.","nordeim\u002Fyoga-studio","nordeim","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F194583642?v=4","01M75CZEV8C3V4DDQNPQ1EA7QJ","01M782NVR9711WMKQXZBPA5P70",0,"MIT","https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002FLICENSE","https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: loop-builder\ndescription: >-\n  Design and scaffold an agent \"loop\" — an unattended, scheduled, self-verifying\n  agent workflow. Use this whenever the user wants to automate a recurring task,\n  schedule an agent, run an agent unattended or overnight, set up monitoring,\n  triage, or alerting, poll something on a cadence, or turn a manual repeated\n  workflow into a self-running one — even if they never say the word \"loop.\" If a\n  request implies \"do this every day \u002F on a schedule \u002F until some condition holds,\n  without me typing each time,\" reach for this skill. It walks the seven-question\n  blueprint, picks the simplest loop pattern, and scaffolds the six building blocks\n  (schedule, isolation, skill, connectors, verifier, state) with a human-gate list\n  and a budget.\n---\n\n# Loop Builder\n\nGuide the user through designing and scaffolding a **loop**: a small self-running\nsystem in which an agent finds work, acts, gets graded by a *separate* checker,\nand repeats until a verifiable condition holds or a budget runs out — without a\nhuman typing each turn.\n\nThe backbone knowledge for everything here is\n`references\u002Floops-and-loop-engineering.md`. It is the source of truth; do not\ncontradict it. If a product mechanic (`\u002Fgoal`, `\u002Floop`, `\u002Fschedule`, dynamic intervals) is\nuncertain, say so and tell the user to verify against current Claude Code \u002F Codex\ndocs — that uncertainty flag is real and must survive into what you generate.\n\n## The one rule that runs through everything\n\n**Durable knowledge → a skill. Changing state → memory.**\n\n- Conventions, commands, \"how we decide done,\" rubrics → the loop's **own**\n  `SKILL.md` (read-only each run).\n- What's been tried, what passed, what's still open → an external **state file**\n  (read *and written* each run).\n\nPutting mutable state inside a `SKILL.md` is the classic anti-pattern — skills are\nversion-controlled and effectively read-only per run. Enforce this split in\neverything you generate. If you catch yourself writing progress into a skill, stop\nand move it to the state file.\n\n## Why loops need this at all\n\nThe agent **starts cold every run** — it forgets everything between runs. So\nconventions, commands, and \"what's already done\" must live outside the context\nwindow, on disk. The agent forgets; the repo does not. Every building block below\nexists to serve that one fact.\n\n## Process\n\nWork in three phases, in order: **elicit → select → scaffold.** Do not jump to\nscaffolding before the seven answers exist — a loop with a missing block is the\nfailure mode, not a shortcut.\n\nCreate a TodoWrite item per phase so nothing is skipped — and a **distinct** item for\nthe Phase 1.5 skill-bank search (dispatch the searcher over `references\u002Fskill-bank\u002F`),\nthe step most easily missed because installed capabilities are already in view and the\nbank is not.\n\n---\n\n## Phase 1 — Elicit the seven decisions\n\nAsk these **one at a time, in order.** Each maps to a building block. Capture the\nanswer before moving on. Don't accept vibes where a predicate is required.\n\n1. **Goal (recursive).** \"What verifiable condition means *done for now*?\" Push for\n   a checkable predicate, not a vibe.\n   - Bad: \"keep the repo healthy.\" Good: \"every P1 issue has an owner and a plan\n     comment.\" If the answer is a vibe, ask \"how would a script know it's true?\"\n2. **Trigger.** \"What fires it — a schedule (cron \u002F `\u002Fschedule` \u002F `\u002Floop`), an\n   event (new PR, inbound email), or run-until-done (`\u002Fgoal \u003Cverifiable condition>`,\n   often paired with `\u002Floop` for a self-terminating loop on a schedule)?\"\n3. **Discovery.** \"How does the agent *find* work each cycle?\" (query the tracker,\n   scan the inbox, diff CI) → a connector.\n4. **Action.** \"What is it allowed to *do*, through which tools?\" → connectors; note\n   whether file work needs an isolated worktree.\n5. **Verification.** \"Who checks the result, and against what? It must be a\n   *separate* checker — a program where one exists.\" → sub-agent or script.\n6. **State \u002F memory.** \"Where does *what's done \u002F what's open* persist outside the\n   context — a markdown ledger, Linear, GitHub issues?\"\n7. **Human gates.** \"Which actions are irreversible or high-blast-radius and need a\n   human approval first — merging, sending external email, spending, deleting?\"\n   This is non-negotiable (see Limitations).\n\nThen capture two more that aren't optional:\n\n- **Knowledge → skill.** \"What conventions should the loop *not* re-derive every\n  run?\" (build\u002Ftest commands, review standards, \"we don't do it that way\").\n- **Budget \u002F stop.** \"What caps a run — max iterations, a token cap, wall-clock?\"\n\nIf the user is fuzzy on the goal, slow down — every other decision hangs off a\ncrisp, checkable goal.\n\n---\n\n## Phase 1.5 — Survey reusable capabilities (don't rebuild what exists)\n\nBefore choosing a pattern, survey what already exists that can *serve a block* instead\nof being built from scratch — the \"compose blocks; don't reach for a framework you\ncan't debug\" principle applied to the build itself. **Search both sources below every\ntime, in order, and map both to the loop's blocks.** Neither is optional.\n\n**1.5a — Installed.** Use `find-skills` (or scan the available skills \u002F MCP list) for\ncapabilities already on the machine (ready to use immediately).\n\n**1.5b — skill-bank — do not skip, even if 1.5a already seems to cover a block.**\n**Search the bank** for capabilities matching the blocks this loop needs. The bank is\nlarge, so don't read it inline — dispatch a search sub-agent:\n\nDispatch a general-purpose sub-agent using the prompt in\n`references\u002Fskill-bank\u002Fsearch-agent.md`, passing the loop's block needs (e.g. \"verifier\nfor X; connector for Y\"). It reads the whole bank — `references\u002Fskill-bank\u002Frecommended.md`\n(curated standouts) and `references\u002Fskill-bank\u002Fcatalog\u002F*.md` (the comprehensive\nlistings) — in its own context, judges relevance semantically, and returns a short\nshortlist per need: the best-fit entries (name, source, block, one-line why, and whether\nit's a recommended standout), or \"none applicable — \u003Creason>\". It prefers standouts when\nthey fit and only reaches into the catalogs when they don't.\n\n*Fallback (if sub-agents aren't available):* read `references\u002Fskill-bank\u002Frecommended.md`\ndirectly, and scan a specific `catalog\u002F\u003Csource>.md` for any block the standouts don't\ncover — same discipline.\n\nBorrow from the shortlist: recommend-and-record only (surface source + install pointer +\na named fallback; never inline external code), confirming license and mechanics against\nsource at borrow time. The point is to surface proven prior art before you settle, so\nyou compose instead of rebuild.\n\nMap what you find — from **both** the installed set (1.5a) and the skill-bank search\nresults (1.5b) — to the blocks the loop needs; recommend only what **genuinely changes\nthe design**, not an exhaustive inventory:\n\n- **Verifier (⑤)** — is there an installed skill or program that can *be* the\n  deterministic check or the evaluator? (e.g. `codex` can act as an independent,\n  *different-model* evaluator — a stronger maker\u002Fchecker split than a same-model\n  sub-agent.)\n- **Connector (④)** — an MCP\u002Fskill that already reads the tracker, the docs, the DB,\n  the inbox? (e.g. `context7` for official library docs; a GitHub MCP for issues.)\n- **Sub-agents (⑤) \u002F workers** — an existing research or worker agent the loop can\n  dispatch rather than hand-rolling one.\n- **Runtime** — is a managed agent runtime available (e.g. **Claude Managed Agents**)\n  that provides scheduling, a grader\u002Fverifier, memory, and approval gates *natively*?\n  If so, consider scaffolding for it instead of hand-rolling those blocks — see\n  `references\u002Fdeploy-claude-managed-agents.md`. It's beta and fast-moving, so verify\n  against current docs and keep a Claude Code fallback.\n\nTwo rules when you recommend reuse, because an external skill is *changing* state you\ndon't control:\n\n1. **Name a fallback** for anything you wire in (\"use `codex`; if unavailable, a\n   general-purpose sub-agent\"), so a cold start still works if the skill is missing\n   or has changed.\n2. **Don't bind to unverified mechanics** — note the skill's name and that its\n   behavior should be confirmed, rather than assuming flags or outputs.\n\nSurface the shortlist to the user and let them choose what to wire in. Record the\nchoices in the `REUSE:` line of the template below.\n\n---\n\n## Phase 2 — Select the simplest fitting pattern\n\nRecommend **one** pattern, then load **only** its reference file (progressive\ndisclosure — don't read all four). Default to ReAct + deterministic verifier and\njustify any escalation.\n\n| If… | Pattern | Load |\n|---|---|---|\n| One workstream; \"done\" is a program-checkable predicate | **ReAct + deterministic verifier** *(default)* | `references\u002Fpattern-react-deterministic-verifier.md` |\n| Clear criteria that need *judgement*, not just a script | **Evaluator–optimizer** | `references\u002Fpattern-evaluator-optimizer.md` |\n| Work genuinely parallelizes into independent subtasks | **Orchestrator–workers** | `references\u002Fpattern-orchestrator-workers.md` |\n| You want a crude baseline \u002F teaching loop | **Ralph** | `references\u002Fpattern-ralph.md` |\n\nGuidance to repeat to the user: **prefer the simplest pattern that works, and\ncompose blocks rather than reaching for a heavy framework you can't debug.** A\nsingle loop with a deterministic verifier beats an elaborate multi-agent system you\ncan't reason about. Escalate to orchestrator–workers only when the work *genuinely*\nparallelizes.\n\n---\n\n## Phase 3 — Emit the template, then scaffold\n\n### 3a. Emit the populated fill-in template\n\nShow the user this template filled with their answers (the literal shape from the\nreference). This is the contract before any files are written:\n\n```\nGOAL (verifiable):      ____\nTRIGGER:                schedule | event | run-until-done → ____\nDISCOVERY (find work):  ____   (connector: ____)\nACTION (do work):       ____   (tools: ____ ; isolation: worktree? y\u002Fn)\nVERIFY (separate check): ____  (deterministic? y\u002Fn)\nREUSE (installed):      ____  (skill\u002FMCP → block it serves ; fallback: ____)\nREUSE (skill-bank):     ____  (bank result → block ; install + fallback ; or \"none applicable — why\")\nSTATE (persist outside): ____  (file | board | issues)\nHUMAN GATES:            ____   (irreversible actions list)\nKNOWLEDGE → skill:      ____   (conventions the loop should not re-derive)\nBUDGET \u002F stop:          ____   (max iterations | token cap | wall-clock)\n```\n\n### 3b. Scaffold the six building blocks\n\nSplit the artifacts by durability so the loop is **invocable by name**, then\nconfirm the locations with the user (these are the defaults):\n\n- **Durable → an installed skill.** Put the loop's own `SKILL.md` (plus its verifier\n  and any templates) under `.claude\u002Fskills\u002F\u003Cloop-name>\u002F`. This matters mechanically:\n  `\u002Fgoal` and `\u002Floop` run a *prompt*, not a folder, so the loop is started by a\n  prompt that **invokes its skill by name\u002Fdescription** (e.g. \"run the `\u003Cloop-name>`\n  loop\"). A `SKILL.md` left under `loops\u002F` is *not* a discoverable skill and can't be\n  invoked that way — it would have to be read by explicit path instead.\n- **Changing → a state folder.** Put run state under `\u003Cproject>\u002Floops\u002F\u003Cloop-name>\u002F`\n  (`STATE.md`, plus any run inputs\u002Foutputs). Read and written every run.\n\nEvery loop gets **all six** blocks; a missing one is what breaks (see the table).\n\n| # | Block | What you write | Durable or changing |\n|---|---|---|---|\n| 1 | **Scheduling** | A trigger stub (cron line \u002F `\u002Fschedule` \u002F `\u002Floop` \u002F `\u002Fgoal \u003Ccondition>` \u002F run-until-verifier) | durable |\n| 2 | **Isolation** | A note\u002Fcommand for git worktrees *if* file work runs in parallel | durable |\n| 3 | **Skill** | The loop's **own** `SKILL.md`, **installed under `.claude\u002Fskills\u002F\u003Cloop-name>\u002F`** — conventions only | **durable** |\n| 4 | **Connectors** | Named MCP\u002Ftools for discovery + action | durable |\n| 5 | **Verifier** | A **separate** check — start from `scripts\u002Fverifier_template.sh` | durable |\n| 6 | **State** | `STATE.md` ledger (or board) in `loops\u002F\u003Cloop-name>\u002F` — what's done \u002F open | **changing** |\n\nConcretely, write:\n\n- **`.claude\u002Fskills\u002F\u003Cloop-name>\u002FSKILL.md`** — the loop's own skill: its goal,\n  conventions, the pattern it uses, how to run it, and what \"done\" means. Installing\n  it here is what makes the loop invocable by a prompt. **Conventions only — never\n  progress.**\n- **`loops\u002F\u003Cloop-name>\u002FSTATE.md`** — a ledger the loop reads and writes each run: a\n  table of items with status, plus \"last run\" notes. This is the changing half; keep\n  it out of the skill.\n- **A verifier script** — copy `scripts\u002Fverifier_template.sh` and wire it to the\n  user's predicate. If it's a \"no P1 unassigned\"-style check, adapt\n  `scripts\u002Fverify_no_p1_unassigned.sh`. The verifier must be *separate* from the\n  generator and deterministic where possible.\n- **A trigger stub** — how the loop is *launched*. `\u002Fgoal` and `\u002Floop` run a\n  *prompt*, so the stub must **name the loop** (`run the \u003Cloop-name> loop`) to load\n  its skill. Match the launcher to the trigger type from Q2 — **don't default\n  everything to `\u002Floop`**, which only makes sense for a recurring cadence:\n    - **Run-until-done \u002F on-demand** (you start it when there's new input — e.g. a\n      comparison loop): just invoke it, `\u002Fgoal \"\u003Ccondition>\"  run the \u003Cloop-name>\n      loop`. **No `\u002Floop`** — it isn't recurring; the trigger is *you*.\n    - **Scheduled** (a cadence): `\u002Fschedule \"\u003Ccron>\" …`, or `\u002Floop \u003Cinterval>` for a\n      simple interval; pair with `\u002Fgoal \"\u003Ccondition>\"` to self-terminate each run.\n    - **Event-driven** (fire when something appears): `\u002Fgoal`\u002F`\u002Floop` can't *listen*\n      for events. Either **(a) poll** — a low-frequency `\u002Floop \u003Cinterval>` that checks\n      a queue\u002Finbox and runs only when there's new work — or **(b) a real external\n      launcher** (a git hook, CI\u002Fwebhook, file-watcher, or a managed runtime; see\n      `references\u002Fdeploy-claude-managed-agents.md`) that invokes the loop when the\n      event fires.\n  `\u002Fgoal` maps directly to the recursive Goal (Q1), so its condition must be the same\n  checkable predicate. **Annotate that these mechanics may have changed — verify\n  against current docs; never fabricate flags.**\n- **`HUMAN-GATES.md`** — the irreversible-actions list **and** the budget\u002Fstop\n  condition, together. This file is mandatory; see below.\n- **Self-reporting (opt-in).** Ask: \"Add self-reporting to this loop? (logs its own\n  errors locally for you to review later — never auto-files)\" If yes, add the hook\n  snippet from `references\u002Ffeedback-to-issue.md` to the verifier. Errors collect\n  locally; the user files them deliberately in a later review session.\n\n### 3c. Always emit human gates + a budget\n\nA loop with no budget and no gates is the failure mode, not the goal. `HUMAN-GATES.md`\nmust contain:\n\n- **Human gates:** every irreversible \u002F high-blast-radius action that requires\n  human approval before execution — merging, sending external email\u002Fmessages,\n  spending money, deleting, publishing. The loop must never cross these\n  autonomously. If the user named none, challenge it: most loops touch at least one.\n- **Budget \u002F stop condition:** max iterations, token cap, or wall-clock — whichever\n  the user chose. Prefer dynamic intervals (short waits while a build finishes, long\n  waits when nothing's pending) where the scheduler supports it.\n\nRefuse to call the loop \"done\" if either is missing.\n\n## Worked examples\n\nThree filled-in templates, chosen to span the grid: different patterns, different\ntrigger types, and the two gate postures (soft \"escalate\" vs. a hard stop on an\nirreversible action). Use them to calibrate the shape of a good answer — don't copy\none verbatim onto a different-shaped problem.\n\n### Example 1 — morning GitHub triage *(ReAct + deterministic · scheduled · soft gate)*\n\n```\nGOAL (verifiable):      zero P1 issues without an assignee AND a plan comment\nTRIGGER:                schedule → every weekday 08:00 (cron \"0 8 * * 1-5\")\nDISCOVERY (find work):  list open P1 issues   (connector: GitHub MCP \u002F gh)\nACTION (do work):       assign an owner, post an initial plan comment (tools: gh)\nVERIFY (separate check): re-query, assert no P1 lacks assignee  (deterministic? y →\n                         scripts\u002Fverify_no_p1_unassigned.sh)\nREUSE (installed):      GitHub MCP → connector (④); fallback: gh CLI\nREUSE (skill-bank):     none applicable — verifier is a bundled deterministic script\n                         (searched the bank; nothing fits better)\nSTATE (persist outside): loops\u002Ftriage\u002FSTATE.md — issues triaged this week\nHUMAN GATES:            none auto-closes; escalate (don't close) anything ambiguous\nKNOWLEDGE → skill:      label taxonomy, what a \"plan comment\" must contain\nBUDGET \u002F stop:          max 25 issues\u002Frun; stop when verifier passes or cap hit\n```\n\nPattern: ReAct + deterministic verifier (one workstream, program-checkable goal).\nComments and labels are reversible, so no hard human gate is required — but the\n*budget* and the *escalate-don't-close* rule still ship.\n\n### Example 2 — inbox draft-replies *(evaluator–optimizer · run-until-done · hard gate)*\n\n```\nGOAL (verifiable):      every unread thread from the last 24h is either drafted-and-\n                        queued or explicitly skipped with a logged reason\nTRIGGER:                run-until-done → you start it when you sit down. No recurring\n                        \u002Floop needed; the loop self-terminates at the goal each run.\nDISCOVERY (find work):  list unread threads in the last 24h  (connector: Gmail MCP)\nACTION (do work):       write a reply into the drafts folder per thread\n                        (tools: Gmail MCP — DRAFTS ONLY, no send scope)\nVERIFY (separate check): a SEPARATE evaluator sub-agent grades each draft against a\n                        rubric (answers the question, right tone, no overcommitment)\n                        — judgement, not a script (deterministic? n)\nREUSE (installed):      Gmail MCP → connector (④); codex as a different-model\n                        evaluator (⑤); fallback: a general-purpose sub-agent\nREUSE (skill-bank):     none applicable — evaluator is the installed different-model\n                        sub-agent (searched the bank)\nSTATE (persist outside): loops\u002Finbox\u002FSTATE.md — threads drafted \u002F skipped + reasons\nHUMAN GATES:            SENDING is a hard gate — the loop only drafts; you press send\nKNOWLEDGE → skill:      voice\u002Ftone guide, what never to commit to in writing,\n                        who to escalate to instead of answering\nBUDGET \u002F stop:          max 20 threads\u002Frun; stop when each is drafted-or-skipped\n```\n\nPattern: Evaluator–optimizer — drafts need *judgement*, so a separate evaluator\ngrades against a rubric (a different-model checker is stronger than a same-model\none). Sending is irreversible and prompt-injection-exposed, so it is a permanent\nhuman gate — the single most important line in this loop. Note the trigger: it is\n*you*, not a clock, so there is no `\u002Floop`.\n\n### Example 3 — nightly CI fix *(ReAct + deterministic · event-driven · isolated · hard gate)*\n\n```\nGOAL (verifiable):      latest nightly build is green, OR a fix PR exists linked to\n                        the failure\nTRIGGER:                event-driven → on nightly-build failure. \u002Fgoal·\u002Floop can't\n                        LISTEN, so either a CI\u002Fwebhook launcher invokes the loop, or a\n                        low-frequency \u002Floop polls build status and runs only on a new\n                        red build.\nDISCOVERY (find work):  read the failing job log + the diff since last green\n                        (connector: GitHub \u002F CI MCP)\nACTION (do work):       reproduce, write a fix, open a PR\n                        (tools: gh ; isolation: worktree? y — parallel file work)\nVERIFY (separate check): re-run build\u002Ftests on the fix branch; green = pass\n                        (deterministic? y — the build itself is the verifier)\nREUSE (installed):      GitHub MCP → connector (④); the build command is the\n                        verifier; fallback: gh CLI\nREUSE (skill-bank):     gstack:investigate (⑤) surfaced by the bank search for root-causing\n                        failures; install: clone into ~\u002F.claude\u002Fskills\u002F; fallback:\n                        inline debugging; verify mechanics against source\nSTATE (persist outside): loops\u002Fci-watch\u002FSTATE.md — failures seen, fix PRs, status\nHUMAN GATES:            opening a PR is fine; MERGING is a hard gate — a human reviews\nKNOWLEDGE → skill:      how to run the build locally, where flaky tests live,\n                        branch \u002F PR conventions\nBUDGET \u002F stop:          max 3 fix attempts\u002Ffailure; stop when green or a PR is open\n```\n\nPattern: ReAct + deterministic verifier again (the build is the program-checkable\npredicate), but event-triggered and worktree-isolated because file work runs in\nparallel. The verifier is the same kind as Example 1 yet the *trigger* and\n*isolation* differ — a reminder that pattern, trigger, and isolation are independent\nchoices.\n\n## Output discipline checklist\n\nBefore declaring the loop scaffolded, confirm:\n\n- [ ] All seven decisions answered; goal is a checkable predicate.\n- [ ] **Installed** capabilities surveyed (1.5a — find-skills \u002F MCP list).\n- [ ] **skill-bank searched** (1.5b): the search sub-agent (or the direct-read fallback) was run over `recommended.md` + `catalog\u002F*.md`; relevant entries surfaced, or \"none applicable — \u003Creason>\" recorded. Not skippable.\n- [ ] Anything wired in (installed or bank) has a named fallback.\n- [ ] One pattern chosen; only its reference was loaded.\n- [ ] Populated template shown to the user.\n- [ ] Six blocks scaffolded as files in the loop folder.\n- [ ] A **separate** verifier exists (script or sub-agent), deterministic if possible.\n- [ ] An **external state file** exists — and no mutable state lives in any SKILL.md.\n- [ ] `HUMAN-GATES.md` lists irreversible actions **and** a budget\u002Fstop condition.\n- [ ] Any `\u002Fgoal` `\u002Floop` `\u002Fschedule` mechanic is flagged \"verify against current docs.\"\n\n## Limitations to bake in (not disclaimers)\n\nState these to the user as the operating posture:\n\n- **Prompt injection is unsolved.** A loop that reads issues, emails, or web content\n  ingests untrusted text every cycle. The durable control is a permanent **human\n  gate on irreversible actions** — never let the loop merge\u002Fsend\u002Fspend\u002Fdelete\n  autonomously.\n- **Verification is the hard part.** Autonomy is only as trustworthy as the checker.\n  Favor deterministic verifiers; keep checker separate from maker; distrust a\n  passing self-grade.\n- **Token economics swing wildly.** Cadence, fan-out, and retries dominate cost. Set\n  explicit budgets and dynamic intervals.\n- **Most of \"agentic\" is plumbing.** The discipline is in the guardrails around the\n  decision, not in the decision being magic.\n\n## Collecting feedback and reporting bugs\n\nBugs and rough edges are captured **locally** — never auto-filed — and filed to\nGitHub under **your own account** only when you explicitly say so.\n\n**Passive capture (on errors).** When loop-builder hits an error or you are clearly\nblocked, append a quiet entry:\n\n```bash\npython3 scripts\u002Ffeedback\u002Fcli.py append --category bug --text \"\u003Cwhat broke + context>\"\n```\n\nThis writes to `~\u002F.loop-builder\u002Ffeedback.jsonl` only. No network call is made.\n\n**Review and file.** When you ask to \"report a bug,\" \"give feedback,\" or \"review\nfeedback,\" load `references\u002Ffeedback-to-issue.md` (on demand — do not read it\nduring normal loop-building). That playbook covers the full flow: `list-open` ->\ncluster -> draft -> dedupe -> sanitize -> **mandatory consent gate** -> `file` ->\n`mark-filed`. Nothing is filed until you see the full title + body + labels and say\nyes.\n\n**Generated loops.** During scaffolding, ask whether to add self-reporting to the\nloop (see the opt-in hook in `references\u002Ffeedback-to-issue.md`). Self-reporting logs\nfailures locally; the user reviews and files them deliberately — the loop never\ncontacts GitHub on its own.\n","# Loop Builder\n\nGuide the user through designing and scaffolding a **loop**: a small self-running\nsystem in which an agent finds work, acts, gets graded by a *separate* checker,\nand repeats until a verifiable condition holds or a budget runs out — without a\nhuman typing each turn.\n\nThe backbone knowledge for everything here is\n`references\u002Floops-and-loop-engineering.md`. It is the source of truth; do not\ncontradict it. If a product mechanic (`\u002Fgoal`, `\u002Floop`, `\u002Fschedule`, dynamic intervals) is\nuncertain, say so and tell the user to verify against current Claude Code \u002F Codex\ndocs — that uncertainty flag is real and must survive into what you generate.\n\n## The one rule that runs through everything\n\n**Durable knowledge → a skill. Changing state → memory.**\n\n- Conventions, commands, \"how we decide done,\" rubrics → the loop's **own**\n  `SKILL.md` (read-only each run).\n- What's been tried, what passed, what's still open → an external **state file**\n  (read *and written* each run).\n\nPutting mutable state inside a `SKILL.md` is the classic anti-pattern — skills are\nversion-controlled and effectively read-only per run. Enforce this split in\neverything you generate. If you catch yourself writing progress into a skill, stop\nand move it to the state file.\n\n## Why loops need this at all\n\nThe agent **starts cold every run** — it forgets everything between runs. So\nconventions, commands, and \"what's already done\" must live outside the context\nwindow, on disk. The agent forgets; the repo does not. Every building block below\nexists to serve that one fact.\n\n## Process\n\nWork in three phases, in order: **elicit → select → scaffold.** Do not jump to\nscaffolding before the seven answers exist — a loop with a missing block is the\nfailure mode, not a shortcut.\n\nCreate a TodoWrite item per phase so nothing is skipped — and a **distinct** item for\nthe Phase 1.5 skill-bank search (dispatch the searcher over `references\u002Fskill-bank\u002F`),\nthe step most easily missed because installed capabilities are already in view and the\nbank is not.\n\n---\n\n## Phase 1 — Elicit the seven decisions\n\nAsk these **one at a time, in order.** Each maps to a building block. Capture the\nanswer before moving on. Don't accept vibes where a predicate is required.\n\n1. **Goal (recursive).** \"What verifiable condition means *done for now*?\" Push for\n   a checkable predicate, not a vibe.\n   - Bad: \"keep the repo healthy.\" Good: \"every P1 issue has an owner and a plan\n     comment.\" If the answer is a vibe, ask \"how would a script know it's true?\"\n2. **Trigger.** \"What fires it — a schedule (cron \u002F `\u002Fschedule` \u002F `\u002Floop`), an\n   event (new PR, inbound email), or run-until-done (`\u002Fgoal \u003Cverifiable condition>`,\n   often paired with `\u002Floop` for a self-terminating loop on a schedule)?\"\n3. **Discovery.** \"How does the agent *find* work each cycle?\" (query the tracker,\n   scan the inbox, diff CI) → a connector.\n4. **Action.** \"What is it allowed to *do*, through which tools?\" → connectors; note\n   whether file work needs an isolated worktree.\n5. **Verification.** \"Who checks the result, and against what? It must be a\n   *separate* checker — a program where one exists.\" → sub-agent or script.\n6. **State \u002F memory.** \"Where does *what's done \u002F what's open* persist outside the\n   context — a markdown ledger, Linear, GitHub issues?\"\n7. **Human gates.** \"Which actions are irreversible or high-blast-radius and need a\n   human approval first — merging, sending external email, spending, deleting?\"\n   This is non-negotiable (see Limitations).\n\nThen capture two more that aren't optional:\n\n- **Knowledge → skill.** \"What conventions should the loop *not* re-derive every\n  run?\" (build\u002Ftest commands, review standards, \"we don't do it that way\").\n- **Budget \u002F stop.** \"What caps a run — max iterations, a token cap, wall-clock?\"\n\nIf the user is fuzzy on the goal, slow down — every other decision hangs off a\ncrisp, checkable goal.\n\n---\n\n## Phase 1.5 — Survey reusable capabilities (don't rebuild what exists)\n\nBefore choosing a pattern, survey what already exists that can *serve a block* instead\nof being built from scratch — the \"compose blocks; don't reach for a framework you\ncan't debug\" principle applied to the build itself. **Search both sources below every\ntime, in order, and map both to the loop's blocks.** Neither is optional.\n\n**1.5a — Installed.** Use `find-skills` (or scan the available skills \u002F MCP list) for\ncapabilities already on the machine (ready to use immediately).\n\n**1.5b — skill-bank — do not skip, even if 1.5a already seems to cover a block.**\n**Search the bank** for capabilities matching the blocks this loop needs. The bank is\nlarge, so don't read it inline — dispatch a search sub-agent:\n\nDispatch a general-purpose sub-agent using the prompt in\n`references\u002Fskill-bank\u002Fsearch-agent.md`, passing the loop's block needs (e.g. \"verifier\nfor X; connector for Y\"). It reads the whole bank — `references\u002Fskill-bank\u002Frecommended.md`\n(curated standouts) and `references\u002Fskill-bank\u002Fcatalog\u002F*.md` (the comprehensive\nlistings) — in its own context, judges relevance semantically, and returns a short\nshortlist per need: the best-fit entries (name, source, block, one-line why, and whether\nit's a recommended standout), or \"none applicable — \u003Creason>\". It prefers standouts when\nthey fit and only reaches into the catalogs when they don't.\n\n*Fallback (if sub-agents aren't available):* read `references\u002Fskill-bank\u002Frecommended.md`\ndirectly, and scan a specific `catalog\u002F\u003Csource>.md` for any block the standouts don't\ncover — same discipline.\n\nBorrow from the shortlist: recommend-and-record only (surface source + install pointer +\na named fallback; never inline external code), confirming license and mechanics against\nsource at borrow time. The point is to surface proven prior art before you settle, so\nyou compose instead of rebuild.\n\nMap what you find — from **both** the installed set (1.5a) and the skill-bank search\nresults (1.5b) — to the blocks the loop needs; recommend only what **genuinely changes\nthe design**, not an exhaustive inventory:\n\n- **Verifier (⑤)** — is there an installed skill or program that can *be* the\n  deterministic check or the evaluator? (e.g. `codex` can act as an independent,\n  *different-model* evaluator — a stronger maker\u002Fchecker split than a same-model\n  sub-agent.)\n- **Connector (④)** — an MCP\u002Fskill that already reads the tracker, the docs, the DB,\n  the inbox? (e.g. `context7` for official library docs; a GitHub MCP for issues.)\n- **Sub-agents (⑤) \u002F workers** — an existing research or worker agent the loop can\n  dispatch rather than hand-rolling one.\n- **Runtime** — is a managed agent runtime available (e.g. **Claude Managed Agents**)\n  that provides scheduling, a grader\u002Fverifier, memory, and approval gates *natively*?\n  If so, consider scaffolding for it instead of hand-rolling those blocks — see\n  `references\u002Fdeploy-claude-managed-agents.md`. It's beta and fast-moving, so verify\n  against current docs and keep a Claude Code fallback.\n\nTwo rules when you recommend reuse, because an external skill is *changing* state you\ndon't control:\n\n1. **Name a fallback** for anything you wire in (\"use `codex`; if unavailable, a\n   general-purpose sub-agent\"), so a cold start still works if the skill is missing\n   or has changed.\n2. **Don't bind to unverified mechanics** — note the skill's name and that its\n   behavior should be confirmed, rather than assuming flags or outputs.\n\nSurface the shortlist to the user and let them choose what to wire in. Record the\nchoices in the `REUSE:` line of the template below.\n\n---\n\n## Phase 2 — Select the simplest fitting pattern\n\nRecommend **one** pattern, then load **only** its reference file (progressive\ndisclosure — don't read all four). Default to ReAct + deterministic verifier and\njustify any escalation.\n\n| If… | Pattern | Load |\n|---|---|---|\n| One workstream; \"done\" is a program-checkable predicate | **ReAct + deterministic verifier** *(default)* | `references\u002Fpattern-react-deterministic-verifier.md` |\n| Clear criteria that need *judgement*, not just a script | **Evaluator–optimizer** | `references\u002Fpattern-evaluator-optimizer.md` |\n| Work genuinely parallelizes into independent subtasks | **Orchestrator–workers** | `references\u002Fpattern-orchestrator-workers.md` |\n| You want a crude baseline \u002F teaching loop | **Ralph** | `references\u002Fpattern-ralph.md` |\n\nGuidance to repeat to the user: **prefer the simplest pattern that works, and\ncompose blocks rather than reaching for a heavy framework you can't debug.** A\nsingle loop with a deterministic verifier beats an elaborate multi-agent system you\ncan't reason about. Escalate to orchestrator–workers only when the work *genuinely*\nparallelizes.\n\n---\n\n## Phase 3 — Emit the template, then scaffold\n\n### 3a. Emit the populated fill-in template\n\nShow the user this template filled with their answers (the literal shape from the\nreference). This is the contract before any files are written:\n\n```\nGOAL (verifiable):      ____\nTRIGGER:                schedule | event | run-until-done → ____\nDISCOVERY (find work):  ____   (connector: ____)\nACTION (do work):       ____   (tools: ____ ; isolation: worktree? y\u002Fn)\nVERIFY (separate check): ____  (deterministic? y\u002Fn)\nREUSE (installed):      ____  (skill\u002FMCP → block it serves ; fallback: ____)\nREUSE (skill-bank):     ____  (bank result → block ; install + fallback ; or \"none applicable — why\")\nSTATE (persist outside): ____  (file | board | issues)\nHUMAN GATES:            ____   (irreversible actions list)\nKNOWLEDGE → skill:      ____   (conventions the loop should not re-derive)\nBUDGET \u002F stop:          ____   (max iterations | token cap | wall-clock)\n```\n\n### 3b. Scaffold the six building blocks\n\nSplit the artifacts by durability so the loop is **invocable by name**, then\nconfirm the locations with the user (these are the defaults):\n\n- **Durable → an installed skill.** Put the loop's own `SKILL.md` (plus its verifier\n  and any templates) under `.claude\u002Fskills\u002F\u003Cloop-name>\u002F`. This matters mechanically:\n  `\u002Fgoal` and `\u002Floop` run a *prompt*, not a folder, so the loop is started by a\n  prompt that **invokes its skill by name\u002Fdescription** (e.g. \"run the `\u003Cloop-name>`\n  loop\"). A `SKILL.md` left under `loops\u002F` is *not* a discoverable skill and can't be\n  invoked that way — it would have to be read by explicit path instead.\n- **Changing → a state folder.** Put run state under `\u003Cproject>\u002Floops\u002F\u003Cloop-name>\u002F`\n  (`STATE.md`, plus any run inputs\u002Foutputs). Read and written every run.\n\nEvery loop gets **all six** blocks; a missing one is what breaks (see the table).\n\n| # | Block | What you write | Durable or changing |\n|---|---|---|---|\n| 1 | **Scheduling** | A trigger stub (cron line \u002F `\u002Fschedule` \u002F `\u002Floop` \u002F `\u002Fgoal \u003Ccondition>` \u002F run-until-verifier) | durable |\n| 2 | **Isolation** | A note\u002Fcommand for git worktrees *if* file work runs in parallel | durable |\n| 3 | **Skill** | The loop's **own** `SKILL.md`, **installed under `.claude\u002Fskills\u002F\u003Cloop-name>\u002F`** — conventions only | **durable** |\n| 4 | **Connectors** | Named MCP\u002Ftools for discovery + action | durable |\n| 5 | **Verifier** | A **separate** check — start from `scripts\u002Fverifier_template.sh` | durable |\n| 6 | **State** | `STATE.md` ledger (or board) in `loops\u002F\u003Cloop-name>\u002F` — what's done \u002F open | **changing** |\n\nConcretely, write:\n\n- **`.claude\u002Fskills\u002F\u003Cloop-name>\u002FSKILL.md`** — the loop's own skill: its goal,\n  conventions, the pattern it uses, how to run it, and what \"done\" means. Installing\n  it here is what makes the loop invocable by a prompt. **Conventions only — never\n  progress.**\n- **`loops\u002F\u003Cloop-name>\u002FSTATE.md`** — a ledger the loop reads and writes each run: a\n  table of items with status, plus \"last run\" notes. This is the changing half; keep\n  it out of the skill.\n- **A verifier script** — copy `scripts\u002Fverifier_template.sh` and wire it to the\n  user's predicate. If it's a \"no P1 unassigned\"-style check, adapt\n  `scripts\u002Fverify_no_p1_unassigned.sh`. The verifier must be *separate* from the\n  generator and deterministic where possible.\n- **A trigger stub** — how the loop is *launched*. `\u002Fgoal` and `\u002Floop` run a\n  *prompt*, so the stub must **name the loop** (`run the \u003Cloop-name> loop`) to load\n  its skill. Match the launcher to the trigger type from Q2 — **don't default\n  everything to `\u002Floop`**, which only makes sense for a recurring cadence:\n    - **Run-until-done \u002F on-demand** (you start it when there's new input — e.g. a\n      comparison loop): just invoke it, `\u002Fgoal \"\u003Ccondition>\"  run the \u003Cloop-name>\n      loop`. **No `\u002Floop`** — it isn't recurring; the trigger is *you*.\n    - **Scheduled** (a cadence): `\u002Fschedule \"\u003Ccron>\" …`, or `\u002Floop \u003Cinterval>` for a\n      simple interval; pair with `\u002Fgoal \"\u003Ccondition>\"` to self-terminate each run.\n    - **Event-driven** (fire when something appears): `\u002Fgoal`\u002F`\u002Floop` can't *listen*\n      for events. Either **(a) poll** — a low-frequency `\u002Floop \u003Cinterval>` that checks\n      a queue\u002Finbox and runs only when there's new work — or **(b) a real external\n      launcher** (a git hook, CI\u002Fwebhook, file-watcher, or a managed runtime; see\n      `references\u002Fdeploy-claude-managed-agents.md`) that invokes the loop when the\n      event fires.\n  `\u002Fgoal` maps directly to the recursive Goal (Q1), so its condition must be the same\n  checkable predicate. **Annotate that these mechanics may have changed — verify\n  against current docs; never fabricate flags.**\n- **`HUMAN-GATES.md`** — the irreversible-actions list **and** the budget\u002Fstop\n  condition, together. This file is mandatory; see below.\n- **Self-reporting (opt-in).** Ask: \"Add self-reporting to this loop? (logs its own\n  errors locally for you to review later — never auto-files)\" If yes, add the hook\n  snippet from `references\u002Ffeedback-to-issue.md` to the verifier. Errors collect\n  locally; the user files them deliberately in a later review session.\n\n### 3c. Always emit human gates + a budget\n\nA loop with no budget and no gates is the failure mode, not the goal. `HUMAN-GATES.md`\nmust contain:\n\n- **Human gates:** every irreversible \u002F high-blast-radius action that requires\n  human approval before execution — merging, sending external email\u002Fmessages,\n  spending money, deleting, publishing. The loop must never cross these\n  autonomously. If the user named none, challenge it: most loops touch at least one.\n- **Budget \u002F stop condition:** max iterations, token cap, or wall-clock — whichever\n  the user chose. Prefer dynamic intervals (short waits while a build finishes, long\n  waits when nothing's pending) where the scheduler supports it.\n\nRefuse to call the loop \"done\" if either is missing.\n\n## Worked examples\n\nThree filled-in templates, chosen to span the grid: different patterns, different\ntrigger types, and the two gate postures (soft \"escalate\" vs. a hard stop on an\nirreversible action). Use them to calibrate the shape of a good answer — don't copy\none verbatim onto a different-shaped problem.\n\n### Example 1 — morning GitHub triage *(ReAct + deterministic · scheduled · soft gate)*\n\n```\nGOAL (verifiable):      zero P1 issues without an assignee AND a plan comment\nTRIGGER:                schedule → every weekday 08:00 (cron \"0 8 * * 1-5\")\nDISCOVERY (find work):  list open P1 issues   (connector: GitHub MCP \u002F gh)\nACTION (do work):       assign an owner, post an initial plan comment (tools: gh)\nVERIFY (separate check): re-query, assert no P1 lacks assignee  (deterministic? y →\n                         scripts\u002Fverify_no_p1_unassigned.sh)\nREUSE (installed):      GitHub MCP → connector (④); fallback: gh CLI\nREUSE (skill-bank):     none applicable — verifier is a bundled deterministic script\n                         (searched the bank; nothing fits better)\nSTATE (persist outside): loops\u002Ftriage\u002FSTATE.md — issues triaged this week\nHUMAN GATES:            none auto-closes; escalate (don't close) anything ambiguous\nKNOWLEDGE → skill:      label taxonomy, what a \"plan comment\" must contain\nBUDGET \u002F stop:          max 25 issues\u002Frun; stop when verifier passes or cap hit\n```\n\nPattern: ReAct + deterministic verifier (one workstream, program-checkable goal).\nComments and labels are reversible, so no hard human gate is required — but the\n*budget* and the *escalate-don't-close* rule still ship.\n\n### Example 2 — inbox draft-replies *(evaluator–optimizer · run-until-done · hard gate)*\n\n```\nGOAL (verifiable):      every unread thread from the last 24h is either drafted-and-\n                        queued or explicitly skipped with a logged reason\nTRIGGER:                run-until-done → you start it when you sit down. No recurring\n                        \u002Floop needed; the loop self-terminates at the goal each run.\nDISCOVERY (find work):  list unread threads in the last 24h  (connector: Gmail MCP)\nACTION (do work):       write a reply into the drafts folder per thread\n                        (tools: Gmail MCP — DRAFTS ONLY, no send scope)\nVERIFY (separate check): a SEPARATE evaluator sub-agent grades each draft against a\n                        rubric (answers the question, right tone, no overcommitment)\n                        — judgement, not a script (deterministic? n)\nREUSE (installed):      Gmail MCP → connector (④); codex as a different-model\n                        evaluator (⑤); fallback: a general-purpose sub-agent\nREUSE (skill-bank):     none applicable — evaluator is the installed different-model\n                        sub-agent (searched the bank)\nSTATE (persist outside): loops\u002Finbox\u002FSTATE.md — threads drafted \u002F skipped + reasons\nHUMAN GATES:            SENDING is a hard gate — the loop only drafts; you press send\nKNOWLEDGE → skill:      voice\u002Ftone guide, what never to commit to in writing,\n                        who to escalate to instead of answering\nBUDGET \u002F stop:          max 20 threads\u002Frun; stop when each is drafted-or-skipped\n```\n\nPattern: Evaluator–optimizer — drafts need *judgement*, so a separate evaluator\ngrades against a rubric (a different-model checker is stronger than a same-model\none). Sending is irreversible and prompt-injection-exposed, so it is a permanent\nhuman gate — the single most important line in this loop. Note the trigger: it is\n*you*, not a clock, so there is no `\u002Floop`.\n\n### Example 3 — nightly CI fix *(ReAct + deterministic · event-driven · isolated · hard gate)*\n\n```\nGOAL (verifiable):      latest nightly build is green, OR a fix PR exists linked to\n                        the failure\nTRIGGER:                event-driven → on nightly-build failure. \u002Fgoal·\u002Floop can't\n                        LISTEN, so either a CI\u002Fwebhook launcher invokes the loop, or a\n                        low-frequency \u002Floop polls build status and runs only on a new\n                        red build.\nDISCOVERY (find work):  read the failing job log + the diff since last green\n                        (connector: GitHub \u002F CI MCP)\nACTION (do work):       reproduce, write a fix, open a PR\n                        (tools: gh ; isolation: worktree? y — parallel file work)\nVERIFY (separate check): re-run build\u002Ftests on the fix branch; green = pass\n                        (deterministic? y — the build itself is the verifier)\nREUSE (installed):      GitHub MCP → connector (④); the build command is the\n                        verifier; fallback: gh CLI\nREUSE (skill-bank):     gstack:investigate (⑤) surfaced by the bank search for root-causing\n                        failures; install: clone into ~\u002F.claude\u002Fskills\u002F; fallback:\n                        inline debugging; verify mechanics against source\nSTATE (persist outside): loops\u002Fci-watch\u002FSTATE.md — failures seen, fix PRs, status\nHUMAN GATES:            opening a PR is fine; MERGING is a hard gate — a human reviews\nKNOWLEDGE → skill:      how to run the build locally, where flaky tests live,\n                        branch \u002F PR conventions\nBUDGET \u002F stop:          max 3 fix attempts\u002Ffailure; stop when green or a PR is open\n```\n\nPattern: ReAct + deterministic verifier again (the build is the program-checkable\npredicate), but event-triggered and worktree-isolated because file work runs in\nparallel. The verifier is the same kind as Example 1 yet the *trigger* and\n*isolation* differ — a reminder that pattern, trigger, and isolation are independent\nchoices.\n\n## Output discipline checklist\n\nBefore declaring the loop scaffolded, confirm:\n\n- [ ] All seven decisions answered; goal is a checkable predicate.\n- [ ] **Installed** capabilities surveyed (1.5a — find-skills \u002F MCP list).\n- [ ] **skill-bank searched** (1.5b): the search sub-agent (or the direct-read fallback) was run over `recommended.md` + `catalog\u002F*.md`; relevant entries surfaced, or \"none applicable — \u003Creason>\" recorded. Not skippable.\n- [ ] Anything wired in (installed or bank) has a named fallback.\n- [ ] One pattern chosen; only its reference was loaded.\n- [ ] Populated template shown to the user.\n- [ ] Six blocks scaffolded as files in the loop folder.\n- [ ] A **separate** verifier exists (script or sub-agent), deterministic if possible.\n- [ ] An **external state file** exists — and no mutable state lives in any SKILL.md.\n- [ ] `HUMAN-GATES.md` lists irreversible actions **and** a budget\u002Fstop condition.\n- [ ] Any `\u002Fgoal` `\u002Floop` `\u002Fschedule` mechanic is flagged \"verify against current docs.\"\n\n## Limitations to bake in (not disclaimers)\n\nState these to the user as the operating posture:\n\n- **Prompt injection is unsolved.** A loop that reads issues, emails, or web content\n  ingests untrusted text every cycle. The durable control is a permanent **human\n  gate on irreversible actions** — never let the loop merge\u002Fsend\u002Fspend\u002Fdelete\n  autonomously.\n- **Verification is the hard part.** Autonomy is only as trustworthy as the checker.\n  Favor deterministic verifiers; keep checker separate from maker; distrust a\n  passing self-grade.\n- **Token economics swing wildly.** Cadence, fan-out, and retries dominate cost. Set\n  explicit budgets and dynamic intervals.\n- **Most of \"agentic\" is plumbing.** The discipline is in the guardrails around the\n  decision, not in the decision being magic.\n\n## Collecting feedback and reporting bugs\n\nBugs and rough edges are captured **locally** — never auto-filed — and filed to\nGitHub under **your own account** only when you explicitly say so.\n\n**Passive capture (on errors).** When loop-builder hits an error or you are clearly\nblocked, append a quiet entry:\n\n```bash\npython3 scripts\u002Ffeedback\u002Fcli.py append --category bug --text \"\u003Cwhat broke + context>\"\n```\n\nThis writes to `~\u002F.loop-builder\u002Ffeedback.jsonl` only. No network call is made.\n\n**Review and file.** When you ask to \"report a bug,\" \"give feedback,\" or \"review\nfeedback,\" load `references\u002Ffeedback-to-issue.md` (on demand — do not read it\nduring normal loop-building). That playbook covers the full flow: `list-open` ->\ncluster -> draft -> dedupe -> sanitize -> **mandatory consent gate** -> `file` ->\n`mark-filed`. Nothing is filed until you see the full title + body + labels and say\nyes.\n\n**Generated loops.** During scaffolding, ask whether to add self-reporting to the\nloop (see the opt-in hook in `references\u002Ffeedback-to-issue.md`). Self-reporting logs\nfailures locally; the user reviews and files them deliberately — the loop never\ncontacts GitHub on its own.\n",{"name":5,"description":7},"main","98424cecddddf49c7a3d5f70f34d8bfff7e74357","skills\u002Floop-builder","https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio","https:\u002F\u002Fgithub.com\u002Fnordeim","A freelance IT engineer","2026-07-06 19:55:45",{"source":8,"skillName":5,"commands":32,"prompt":42,"zipAvailable":43},[33,36,39],{"tool":34,"command":35},"npx","npx skills add nordeim\u002Fyoga-studio --skill loop-builder",{"tool":37,"command":38},"bunx","bunx skills add nordeim\u002Fyoga-studio --skill loop-builder",{"tool":40,"command":41},"pnpm","pnpm dlx skills add nordeim\u002Fyoga-studio --skill loop-builder","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002FSKILL.md",true,[45,51,55,59,63,67,73,77,81,85,89,93,97,101,105,109,113,117,121,125,129,133,137,141,145,149,153,157,161,165,169,173,177,181,185,189,193,197,201,207,211,215,219,223,227,231,235,239,243,247,251,255,259,263,267,271,275,279,283,287,291,295,299,302,306,310,314,318,322],{"path":46,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":49,"sourceHtmlUrl":50},".github\u002FISSUE_TEMPLATE\u002Fbug_report.yml","REFERENCE","参考文档",622,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002F.github\u002FISSUE_TEMPLATE\u002Fbug_report.yml",{"path":52,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":53,"sourceHtmlUrl":54},".github\u002FISSUE_TEMPLATE\u002Fconfig.yml",27,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002F.github\u002FISSUE_TEMPLATE\u002Fconfig.yml",{"path":56,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":57,"sourceHtmlUrl":58},".github\u002FISSUE_TEMPLATE\u002Ffeedback.yml",286,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002F.github\u002FISSUE_TEMPLATE\u002Ffeedback.yml",{"path":60,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":61,"sourceHtmlUrl":62},".github\u002Fworkflows\u002Fsecret-scan.yml",1581,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002F.github\u002Fworkflows\u002Fsecret-scan.yml",{"path":64,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":65,"sourceHtmlUrl":66},".github\u002Fworkflows\u002Ftests.yml",776,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002F.github\u002Fworkflows\u002Ftests.yml",{"path":68,"resourceType":69,"resourceTypeDesc":70,"sizeBytes":71,"sourceHtmlUrl":72},".gitignore","OTHER","其他",577,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002F.gitignore",{"path":74,"resourceType":69,"resourceTypeDesc":70,"sizeBytes":75,"sourceHtmlUrl":76},".gitleaks.toml",870,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002F.gitleaks.toml",{"path":78,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":79,"sourceHtmlUrl":80},".pre-commit-config.yaml",1354,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002F.pre-commit-config.yaml",{"path":82,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":83,"sourceHtmlUrl":84},"AGENTS.md",1665,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002FAGENTS.md",{"path":86,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":87,"sourceHtmlUrl":88},"docs\u002Fdesign-spec.md",4163,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Fdocs\u002Fdesign-spec.md",{"path":90,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":91,"sourceHtmlUrl":92},"docs\u002Fsuperpowers\u002Fplans\u002F2026-06-19-skill-bank.md",26826,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Fdocs\u002Fsuperpowers\u002Fplans\u002F2026-06-19-skill-bank.md",{"path":94,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":95,"sourceHtmlUrl":96},"docs\u002Fsuperpowers\u002Fplans\u002F2026-06-20-catalog-multipath-ref.md",14283,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Fdocs\u002Fsuperpowers\u002Fplans\u002F2026-06-20-catalog-multipath-ref.md",{"path":98,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":99,"sourceHtmlUrl":100},"docs\u002Fsuperpowers\u002Fplans\u002F2026-06-20-skill-bank-semantic-search.md",22046,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Fdocs\u002Fsuperpowers\u002Fplans\u002F2026-06-20-skill-bank-semantic-search.md",{"path":102,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":103,"sourceHtmlUrl":104},"docs\u002Fsuperpowers\u002Fplans\u002F2026-06-20-skill-bank-two-tier.md",27690,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Fdocs\u002Fsuperpowers\u002Fplans\u002F2026-06-20-skill-bank-two-tier.md",{"path":106,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":107,"sourceHtmlUrl":108},"docs\u002Fsuperpowers\u002Fplans\u002F2026-06-22-feedback-to-issue.md",39402,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Fdocs\u002Fsuperpowers\u002Fplans\u002F2026-06-22-feedback-to-issue.md",{"path":110,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":111,"sourceHtmlUrl":112},"docs\u002Fsuperpowers\u002Fspecs\u002F2026-06-19-skill-bank-design.md",8754,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Fdocs\u002Fsuperpowers\u002Fspecs\u002F2026-06-19-skill-bank-design.md",{"path":114,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":115,"sourceHtmlUrl":116},"docs\u002Fsuperpowers\u002Fspecs\u002F2026-06-20-catalog-multipath-ref-design.md",5146,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Fdocs\u002Fsuperpowers\u002Fspecs\u002F2026-06-20-catalog-multipath-ref-design.md",{"path":118,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":119,"sourceHtmlUrl":120},"docs\u002Fsuperpowers\u002Fspecs\u002F2026-06-20-skill-bank-semantic-search-design.md",7841,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Fdocs\u002Fsuperpowers\u002Fspecs\u002F2026-06-20-skill-bank-semantic-search-design.md",{"path":122,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":123,"sourceHtmlUrl":124},"docs\u002Fsuperpowers\u002Fspecs\u002F2026-06-20-skill-bank-two-tier-design.md",8216,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Fdocs\u002Fsuperpowers\u002Fspecs\u002F2026-06-20-skill-bank-two-tier-design.md",{"path":126,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":127,"sourceHtmlUrl":128},"docs\u002Fsuperpowers\u002Fspecs\u002F2026-06-22-feedback-to-issue-design.md",9998,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Fdocs\u002Fsuperpowers\u002Fspecs\u002F2026-06-22-feedback-to-issue-design.md",{"path":130,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":131,"sourceHtmlUrl":132},"evals\u002Fevals.json",7408,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Fevals\u002Fevals.json",{"path":134,"resourceType":134,"resourceTypeDesc":135,"sizeBytes":136,"sourceHtmlUrl":15},"LICENSE","许可证",1065,{"path":138,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":139,"sourceHtmlUrl":140},"README.md",22763,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002FREADME.md",{"path":142,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":143,"sourceHtmlUrl":144},"references\u002Fdeploy-claude-managed-agents.md",4645,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Freferences\u002Fdeploy-claude-managed-agents.md",{"path":146,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":147,"sourceHtmlUrl":148},"references\u002Ffeedback-to-issue.md",8106,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Freferences\u002Ffeedback-to-issue.md",{"path":150,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":151,"sourceHtmlUrl":152},"references\u002Floops-and-loop-engineering.md",18430,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Freferences\u002Floops-and-loop-engineering.md",{"path":154,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":155,"sourceHtmlUrl":156},"references\u002Fpattern-evaluator-optimizer.md",2541,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Freferences\u002Fpattern-evaluator-optimizer.md",{"path":158,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":159,"sourceHtmlUrl":160},"references\u002Fpattern-orchestrator-workers.md",2852,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Freferences\u002Fpattern-orchestrator-workers.md",{"path":162,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":163,"sourceHtmlUrl":164},"references\u002Fpattern-ralph.md",2399,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Freferences\u002Fpattern-ralph.md",{"path":166,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":167,"sourceHtmlUrl":168},"references\u002Fpattern-react-deterministic-verifier.md",2700,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Freferences\u002Fpattern-react-deterministic-verifier.md",{"path":170,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":171,"sourceHtmlUrl":172},"references\u002Fskill-bank\u002Fcatalog\u002Fawesome-claude-skills.md",8085,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Freferences\u002Fskill-bank\u002Fcatalog\u002Fawesome-claude-skills.md",{"path":174,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":175,"sourceHtmlUrl":176},"references\u002Fskill-bank\u002Fcatalog\u002Fclaude-for-legal.md",44826,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Freferences\u002Fskill-bank\u002Fcatalog\u002Fclaude-for-legal.md",{"path":178,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":179,"sourceHtmlUrl":180},"references\u002Fskill-bank\u002Fcatalog\u002Fecc.md",68618,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Freferences\u002Fskill-bank\u002Fcatalog\u002Fecc.md",{"path":182,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":183,"sourceHtmlUrl":184},"references\u002Fskill-bank\u002Fcatalog\u002Fgstack.md",5381,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Freferences\u002Fskill-bank\u002Fcatalog\u002Fgstack.md",{"path":186,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":187,"sourceHtmlUrl":188},"references\u002Fskill-bank\u002Fcatalog\u002Fsuperpowers.md",2773,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Freferences\u002Fskill-bank\u002Fcatalog\u002Fsuperpowers.md",{"path":190,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":191,"sourceHtmlUrl":192},"references\u002Fskill-bank\u002Frecommended.md",5941,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Freferences\u002Fskill-bank\u002Frecommended.md",{"path":194,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":195,"sourceHtmlUrl":196},"references\u002Fskill-bank\u002Fsearch-agent.md",1820,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Freferences\u002Fskill-bank\u002Fsearch-agent.md",{"path":198,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":199,"sourceHtmlUrl":200},"references\u002Fskill-bank\u002Fsources.yml",2020,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Freferences\u002Fskill-bank\u002Fsources.yml",{"path":202,"resourceType":203,"resourceTypeDesc":204,"sizeBytes":205,"sourceHtmlUrl":206},"scripts\u002Fbuild_skill_bank_catalog.sh","SCRIPT","脚本",3214,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Fscripts\u002Fbuild_skill_bank_catalog.sh",{"path":208,"resourceType":203,"resourceTypeDesc":204,"sizeBytes":209,"sourceHtmlUrl":210},"scripts\u002Ffeedback\u002Fcli.py",2504,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Fscripts\u002Ffeedback\u002Fcli.py",{"path":212,"resourceType":203,"resourceTypeDesc":204,"sizeBytes":213,"sourceHtmlUrl":214},"scripts\u002Ffeedback\u002Fdedupe.py",1798,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Fscripts\u002Ffeedback\u002Fdedupe.py",{"path":216,"resourceType":203,"resourceTypeDesc":204,"sizeBytes":217,"sourceHtmlUrl":218},"scripts\u002Ffeedback\u002Ffeedback_log.py",2138,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Fscripts\u002Ffeedback\u002Ffeedback_log.py",{"path":220,"resourceType":203,"resourceTypeDesc":204,"sizeBytes":221,"sourceHtmlUrl":222},"scripts\u002Ffeedback\u002Ffile_issue.py",3852,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Fscripts\u002Ffeedback\u002Ffile_issue.py",{"path":224,"resourceType":203,"resourceTypeDesc":204,"sizeBytes":225,"sourceHtmlUrl":226},"scripts\u002Ffeedback\u002Fsanitize.py",1516,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Fscripts\u002Ffeedback\u002Fsanitize.py",{"path":228,"resourceType":203,"resourceTypeDesc":204,"sizeBytes":229,"sourceHtmlUrl":230},"scripts\u002Fformat_catalog.sh",1730,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Fscripts\u002Fformat_catalog.sh",{"path":232,"resourceType":203,"resourceTypeDesc":204,"sizeBytes":233,"sourceHtmlUrl":234},"scripts\u002Flint_skill_bank_catalog.sh",1674,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Fscripts\u002Flint_skill_bank_catalog.sh",{"path":236,"resourceType":203,"resourceTypeDesc":204,"sizeBytes":237,"sourceHtmlUrl":238},"scripts\u002Flint_skill_bank_recommended.sh",2790,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Fscripts\u002Flint_skill_bank_recommended.sh",{"path":240,"resourceType":203,"resourceTypeDesc":204,"sizeBytes":241,"sourceHtmlUrl":242},"scripts\u002Frefresh_skill_bank.sh",3920,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Fscripts\u002Frefresh_skill_bank.sh",{"path":244,"resourceType":203,"resourceTypeDesc":204,"sizeBytes":245,"sourceHtmlUrl":246},"scripts\u002Ftests\u002Ffeedback\u002Ftest_cli.py",1946,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Fscripts\u002Ftests\u002Ffeedback\u002Ftest_cli.py",{"path":248,"resourceType":203,"resourceTypeDesc":204,"sizeBytes":249,"sourceHtmlUrl":250},"scripts\u002Ftests\u002Ffeedback\u002Ftest_dedupe.py",1909,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Fscripts\u002Ftests\u002Ffeedback\u002Ftest_dedupe.py",{"path":252,"resourceType":203,"resourceTypeDesc":204,"sizeBytes":253,"sourceHtmlUrl":254},"scripts\u002Ftests\u002Ffeedback\u002Ftest_feedback_log.py",1700,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Fscripts\u002Ftests\u002Ffeedback\u002Ftest_feedback_log.py",{"path":256,"resourceType":203,"resourceTypeDesc":204,"sizeBytes":257,"sourceHtmlUrl":258},"scripts\u002Ftests\u002Ffeedback\u002Ftest_file_issue.py",2163,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Fscripts\u002Ftests\u002Ffeedback\u002Ftest_file_issue.py",{"path":260,"resourceType":203,"resourceTypeDesc":204,"sizeBytes":261,"sourceHtmlUrl":262},"scripts\u002Ftests\u002Ffeedback\u002Ftest_sanitize.py",1627,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Fscripts\u002Ftests\u002Ffeedback\u002Ftest_sanitize.py",{"path":264,"resourceType":203,"resourceTypeDesc":204,"sizeBytes":265,"sourceHtmlUrl":266},"scripts\u002Ftests\u002Ffixtures\u002Fp1_failing.json",280,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Fscripts\u002Ftests\u002Ffixtures\u002Fp1_failing.json",{"path":268,"resourceType":203,"resourceTypeDesc":204,"sizeBytes":269,"sourceHtmlUrl":270},"scripts\u002Ftests\u002Ffixtures\u002Fp1_passing.json",281,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Fscripts\u002Ftests\u002Ffixtures\u002Fp1_passing.json",{"path":272,"resourceType":203,"resourceTypeDesc":204,"sizeBytes":273,"sourceHtmlUrl":274},"scripts\u002Ftests\u002Ffixtures\u002Fskill_bank\u002Fcatalog_bad.md",252,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Fscripts\u002Ftests\u002Ffixtures\u002Fskill_bank\u002Fcatalog_bad.md",{"path":276,"resourceType":203,"resourceTypeDesc":204,"sizeBytes":277,"sourceHtmlUrl":278},"scripts\u002Ftests\u002Ffixtures\u002Fskill_bank\u002Fcatalog_expected.md",298,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Fscripts\u002Ftests\u002Ffixtures\u002Fskill_bank\u002Fcatalog_expected.md",{"path":280,"resourceType":203,"resourceTypeDesc":204,"sizeBytes":281,"sourceHtmlUrl":282},"scripts\u002Ftests\u002Ffixtures\u002Fskill_bank\u002Fcatalog_good.md",271,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Fscripts\u002Ftests\u002Ffixtures\u002Fskill_bank\u002Fcatalog_good.md",{"path":284,"resourceType":203,"resourceTypeDesc":204,"sizeBytes":285,"sourceHtmlUrl":286},"scripts\u002Ftests\u002Ffixtures\u002Fskill_bank\u002Findex_bad.md",370,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Fscripts\u002Ftests\u002Ffixtures\u002Fskill_bank\u002Findex_bad.md",{"path":288,"resourceType":203,"resourceTypeDesc":204,"sizeBytes":289,"sourceHtmlUrl":290},"scripts\u002Ftests\u002Ffixtures\u002Fskill_bank\u002Findex_good.md",513,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Fscripts\u002Ftests\u002Ffixtures\u002Fskill_bank\u002Findex_good.md",{"path":292,"resourceType":203,"resourceTypeDesc":204,"sizeBytes":293,"sourceHtmlUrl":294},"scripts\u002Ftests\u002Ffixtures\u002Fskill_bank\u002Fupstream_drift.txt",28,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Fscripts\u002Ftests\u002Ffixtures\u002Fskill_bank\u002Fupstream_drift.txt",{"path":296,"resourceType":203,"resourceTypeDesc":204,"sizeBytes":297,"sourceHtmlUrl":298},"scripts\u002Ftests\u002Ffixtures\u002Fskill_bank\u002Fupstream_empty.txt",19,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Fscripts\u002Ftests\u002Ffixtures\u002Fskill_bank\u002Fupstream_empty.txt",{"path":300,"resourceType":203,"resourceTypeDesc":204,"sizeBytes":53,"sourceHtmlUrl":301},"scripts\u002Ftests\u002Ffixtures\u002Fskill_bank\u002Fupstream_insync.txt","https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Fscripts\u002Ftests\u002Ffixtures\u002Fskill_bank\u002Fupstream_insync.txt",{"path":303,"resourceType":203,"resourceTypeDesc":204,"sizeBytes":304,"sourceHtmlUrl":305},"scripts\u002Ftests\u002Ffixtures\u002Fskill_bank\u002Fupstream_stale.txt",41,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Fscripts\u002Ftests\u002Ffixtures\u002Fskill_bank\u002Fupstream_stale.txt",{"path":307,"resourceType":203,"resourceTypeDesc":204,"sizeBytes":308,"sourceHtmlUrl":309},"scripts\u002Ftests\u002Ftest_feedback.sh",342,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Fscripts\u002Ftests\u002Ftest_feedback.sh",{"path":311,"resourceType":203,"resourceTypeDesc":204,"sizeBytes":312,"sourceHtmlUrl":313},"scripts\u002Ftests\u002Ftest_skill_bank.sh",3562,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Fscripts\u002Ftests\u002Ftest_skill_bank.sh",{"path":315,"resourceType":203,"resourceTypeDesc":204,"sizeBytes":316,"sourceHtmlUrl":317},"scripts\u002Ftests\u002Ftest_verifiers.sh",1433,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Fscripts\u002Ftests\u002Ftest_verifiers.sh",{"path":319,"resourceType":203,"resourceTypeDesc":204,"sizeBytes":320,"sourceHtmlUrl":321},"scripts\u002Fverifier_template.sh",1380,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Fscripts\u002Fverifier_template.sh",{"path":323,"resourceType":203,"resourceTypeDesc":204,"sizeBytes":324,"sourceHtmlUrl":325},"scripts\u002Fverify_no_p1_unassigned.sh",2029,"https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Floop-builder\u002Fscripts\u002Fverify_no_p1_unassigned.sh",[327,336,344,352,359,367],{"skillNo":328,"name":329,"nameZh":6,"description":330,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":331,"licenseFileUrl":332,"sourceHtmlUrl":333,"categories":334,"keywords":335,"publishedAt":19},"01M783VZ3397F6J78J4FTDJN4S","webapp-testing","Web application testing principles. E2E, Playwright, deep audit strategies.","Apache-2.0","https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Fwebapp-testing\u002FLICENSE.txt","https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Fwebapp-testing\u002FSKILL.md",[],[],{"skillNo":337,"name":338,"nameZh":6,"description":339,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":340,"sourceHtmlUrl":341,"categories":342,"keywords":343,"publishedAt":19},"01M783VE24VSNZVBGZD1S88XCF","web-search","Implement web search capabilities using the z-ai-web-dev-sdk. Use this skill when the user needs to search for real-time information from the web, retrieve up-to-date content beyond the knowledge cutoff, or find the latest news and data. Returns structured search results with URLs, snippets, and metadata.","https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Fweb-search\u002FLICENSE.txt","https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Fweb-search\u002FSKILL.md",[],[],{"skillNo":345,"name":346,"nameZh":6,"description":347,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":348,"sourceHtmlUrl":349,"categories":350,"keywords":351,"publishedAt":19},"01M783V8S6X871D5Z7X8TJDCQS","web-reader","Implement web page content extraction capabilities using the z-ai-web-dev-sdk. Use this skill when the user needs to scrape web pages, extract article content, retrieve page metadata, or build applications that process web content. Supports automatic content extraction with title, HTML, and publication time retrieval.","https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Fweb-reader\u002FLICENSE.txt","https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Fweb-reader\u002FSKILL.md",[],[],{"skillNo":353,"name":354,"nameZh":6,"description":355,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":356,"categories":357,"keywords":358,"publishedAt":19},"01M783V2QN3TQBPB06WK5C3PED","web-frameworks","Build modern full-stack web applications with Next.js 16 + React 19 (App Router, Server Components, RSC, PPR, cacheComponents, async params, SSR, SSG, ISR, standalone output), Turborepo (monorepo, task pipelines, remote caching, parallel execution), and RemixIcon (3100+ SVG icons). Use when creating React applications, implementing server-side rendering, setting up monorepos, optimizing build performance and caching, or working with TypeScript full-stack projects.","https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Fweb-frameworks\u002FSKILL.md",[],[],{"skillNo":360,"name":361,"nameZh":6,"description":362,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":363,"sourceHtmlUrl":364,"categories":365,"keywords":366,"publishedAt":19},"01M783TJF06W3R5KVKWH4AGVE5","video-understand","Implement specialized video understanding capabilities using the z-ai-web-dev-sdk. Use this skill when the user needs to analyze video content, understand motion and temporal sequences, extract information from video frames, describe video scenes, or perform video-based AI analysis. Optimized for MP4, AVI, MOV, and other common video formats.","https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Fvideo-understand\u002FLICENSE.txt","https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Fvideo-understand\u002FSKILL.md",[],[],{"skillNo":368,"name":369,"nameZh":6,"description":370,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":371,"sourceHtmlUrl":372,"categories":373,"keywords":374,"publishedAt":19},"01M783TC8J5JMJY4369A7HCJ3G","video-generation","Implement AI-powered video generation capabilities using the z-ai-web-dev-sdk. Use this skill when the user needs to generate videos from text prompts or images, create video content programmatically, or build applications that produce video outputs. Supports asynchronous task management with status polling and result retrieval.","https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Fvideo-generation\u002FLICENSE.txt","https:\u002F\u002Fgithub.com\u002Fnordeim\u002Fyoga-studio\u002Fblob\u002Fmain\u002Fskills\u002Fvideo-generation\u002FSKILL.md",[],[]]