spec-kitty.specify
Create a mission specification
- License 许可
- MIT
- 最近更新
- 2026-11-15
- Stars
- ★ 2
- Forks
- 0
01 前言 Front Matter
- name
- spec-kitty.specify
- description
- Create a mission specification
- user-invocable
- true
03 SKILL.md 正文
Startup Upgrade Check
Run this at most once per active agent session before the first Spec Kitty command workflow.
If you already ran spec-kitty upgrade --agent-check --json in this session, reuse that result and skip this block.
Do not run or announce an upgrade check again for later Spec Kitty commands in the same session.
Otherwise, before continuing, run:
spec-kitty upgrade --agent-check --json
If JSON action is none, continue.
If action is auto_upgrade, run upgrade_command before continuing. If it fails, tell the user and continue with the current Spec Kitty version.
If action is guidance, show upgrade_note briefly, then continue.
If action is prompt, ask the user with the host-native question UI when available:
Spec Kitty {latest_version} is available. You are on {installed_version}. Upgrade now?
Use these choices:
- Upgrade now (recommended) - record
upgrade_now, runupgrade_command, then continue. - Always keep me up to date - record
always, runupgrade_command, then continue. - Not now - record
not_now, then continue. - Never ask again - record
never_ask, then continue.
Record the selected choice before continuing:
spec-kitty upgrade --agent-choice <upgrade_now|always|not_now|never_ask> --agent-latest <latest_version> --json
If no host-native question UI is available, present the same four choices in plain text and wait for the user.
In non-interactive hosts, choose not_now and continue.
/spec-kitty.specify - Create Mission Specification
Version: 0.11.0+
REASONS Guidance — Specify
This project's active charter includes the SPDD/REASONS artifacts. While capturing the spec, populate or update these REASONS canvas sections:
- Requirements — problem statement, acceptance criteria, definition of done.
- Entities — domain concepts, relationships, canonical glossary terms.
Reference: kitty-specs/<mission>/reasons-canvas.md if present. If the canvas
does not yet exist, author it with the section headings listed above (see the
reasons-canvas-writing styleguide).
Charter directives take precedence over canvas content.
📍 WORKING DIRECTORY: Stay in the repository root checkout
IMPORTANT: Specify works in the repository root checkout. NO worktrees are created.
# Run from the repository root checkout:
cd /path/to/project/root # Your repository root checkout
# All planning artifacts are created in the project root and committed:
# - kitty-specs/<mission_slug>/spec.md → Created in project root
# (use the mission_slug returned by `mission create`; the numeric NNN- prefix
# is display-only and is assigned at merge time)
# - Committed to target branch (from create JSON: target_branch/base_branch)
# - NO worktrees created
Worktrees are created later during /spec-kitty.implement, after task finalization computes execution lanes.
Mission Handle Rule
Create the Mission scaffold before asking any discovery or brief-intake question. The initial invocation text is enough to derive a provisional identity; it is not the final specification. Mission creation establishes the handle and empty scaffold only: it does not authorize writing substantive spec content or committing it.
- Do not pass
--missiontospec-kitty agent mission branch-contextor to the initialspec-kitty agent mission createcommand. - Derive a concise provisional title and kebab-case slug from the initial user
input, resolve branch intent, then run
mission createbefore interview. - After
createsucceeds, use the returnedmission_slugormission_idas<handle>for every Decision Moment and command that operates on the Mission. <handle>can be the mission'smission_id(ULID),mid8(first 8 chars of the ULID), ormission_slug.- The resolver disambiguates by
mission_idand returns a structuredMISSION_AMBIGUOUS_SELECTORerror on ambiguity — there is no silent fallback.
User Input
The content of the user's message that invoked this skill (everything after the skill invocation token, e.g. after /spec-kitty.<command> or $spec-kitty.<command>) is the User Input referenced elsewhere in these instructions.
You MUST consider this user input before proceeding (if not empty).
Mission Type and Creation Metadata Bootstrap (before create)
Mission creation emits the canonical MissionCreated lifecycle event and
selects the type-specific spec scaffold. Resolve and freeze the activated
Mission type before mission create. Mission type cannot be changed after
creation.
-
List the Mission types activated for this project:
spec-kitty mission list --json -
Infer the type from the initial invocation text, an available brief, or an explicit user selection. Building or changing software normally selects
software-dev; investigation or analysis may selectresearch. Never select a type that is absent from the activation-filtered list. -
If the type is genuinely ambiguous, ask one short Mission-type selection question before
create. This is immutable lifecycle bootstrap and operational preflight, not a discovery interview; do not open a Decision Moment for it. If the user supplied an explicit type, do not ask again. -
Derive and freeze a truthful creation snapshot:
friendly_name,purpose_tldr, andpurpose_context. These values describe the initial Mission purpose recorded byMissionCreated; the later confirmed Intent Summary governsspec.mdand may be more precise.
This prompt owns only the software-dev specification contract. It may
bootstrap another activated Mission type so the Mission exists before its
agent-host interaction, but after create/resume it must hand off immediately to
that type's runtime action as described below. Never apply this prompt's
FR/NFR/C schema, software quality checklist, spec-commit, or setup-plan
steps to a non-software-dev Mission.
Do not rewrite friendly_name, purpose_tldr, purpose_context, or
mission_type during specify. There is no metadata-update lifecycle event in
this flow, so mutating those fields after creation would make meta.json
disagree with canonical event history.
Execution Order Contract
Follow these transitions in order; do not advance by merely finding similarly named headings elsewhere in this prompt:
- Load charter/brief context, resolve branch strategy, freeze the Mission type and creation metadata, then perform the resume probe.
- If no matching scaffold exists, run
spec-kitty agent mission createexactly once with the frozen type and metadata. If a matching incomplete scaffold exists, reuse it and do not create another Mission. - After creation or verified resume succeeds, branch on the frozen type:
- For
software-dev, runspec-kitty agent decision openfor the first interview Decision Moment. - For any other type, first query its runtime state with
spec-kitty next --agent <agent> --mission <handle> --json. Only when that query reportskind: "query"andmission_state: "not_started"may you issue the first runtime action with--result success. If the query reports an existing run or outstanding step, never report success for work you did not execute; stop this prompt and resume or recover that type-specific action explicitly.
- For
- For
software-dev, ask the discovery question. Resolve, defer, or cancel that Decision Moment before opening the next one. - Only after the user confirms the Intent Summary, write substantive
requirements and run
spec-kitty spec-commitwith bothspec.mdandmeta.json.
Primary Invariant: What Are We Building?
This workflow answers "What are we building?" before it writes substantive artifacts. The raw invocation text is only a starting point for discovery, not the final truth.
Before writing substantive spec.md content or committing it, you MUST have
one of these:
- A completed discovery interview with an acknowledged Intent Summary.
- A brief-intake summary and extracted requirement set explicitly confirmed by the user.
- An explicit user instruction to minimize or skip discovery; even then, record the minimal confirmed scenario and assumptions in the Intent Summary.
For non-trivial work, the confirmed Intent Summary must cover the primary actor, trigger, desired outcome, one rule or invariant, and any canonical domain term or boundary that materially affects the work.
The early Mission scaffold exists so the interview can record Decision Moments; it does not weaken this confirmed-intent gate.
Branch Strategy Confirmation (MANDATORY)
Branch target and strategy confirmation is operational preflight, not a
discovery interview. It may require a user response before create, but do
not open a Decision Moment for it. Do not ask a product, requirements, or
implementation question before create succeeds. For an empty invocation,
one bootstrap identity prompt is permitted before create solely to obtain a
working Mission name; it is not a discovery question. The Decision Moment
Protocol begins only after create succeeds.
Before discovery, resolve branch intent through the Python helper, not by probing git directly:
spec-kitty agent mission branch-context --json
If the user already told you the intended landing branch, pass it explicitly:
spec-kitty agent mission branch-context --json --target-branch <intended-branch>
Parse the JSON and, in your next reply, explicitly tell the user:
- Current branch at workflow start:
current_branch - Default planning/base branch if you create the mission right now:
planning_base_branch - Final merge target for completed changes:
merge_target_branch - Whether
branch_matches_targetis true or false - If that is not the intended landing branch, stop and ask which branch should receive this mission before you run
create
Never talk generically about main or "the default branch". Name the actual branch values from the helper JSON. Do not shell out to git to resolve branch state for this prompt — the helper is the source of truth.
Primary-branch recommendation (issue #765)
The helper JSON also returns a primary-branch recommendation payload:
primary_branch— the repository's primary branch (e.g.main)current_is_primary—truewhen you are standing on that primary branchrecommended_strategy—feature-branch(start a dedicated branch) orstayreason— a human-readable explanation you should relay to the user
When current_is_primary is true, you must have an explicit branching-strategy conversation before calling create and create on a dedicated PR branch:
-
Relay the
reasonto the user and ask whether they expect to open a pull request for this work later (the default assumption for mission work is yes). -
If they expect a PR (recommended path): recommend starting on a dedicated PR branch now, and propose a name derived from the provisional slug — e.g.
feat/<slug>(usefix/<slug>for a bug-fix mission). Pass that branch tocreatewith--start-branchso the CLI creates/switches to it before writing any mission artifacts:spec-kitty agent mission create "<slug>" \ --mission-type "<mission-type>" \ --friendly-name "<title>" \ --purpose-tldr "<purpose_tldr>" \ --purpose-context "<purpose_context>" \ --json \ --pr-bound \ --branch-strategy already-confirmed \ --start-branch feat/<slug>Use the full
createcommand in the Outline section below; the example here only shows the required branch flags. Do not run a separate rawgit switchfor this flow. -
If they do not expect a PR: a dedicated non-protected topic (PR) branch is still required for specify. Planning artifacts do not fall back to the coordination branch, and
spec-commitrefuses a protected primary ref. Explain that invariant, propose a non-protectedfeat/<slug>orfix/<slug>branch, and use--start-branch. If the user declines, stop beforecreaterather than beginning a Mission that cannot complete specify safely.
When current_is_primary is false, you are already on a non-primary branch — no branch switch is needed; proceed normally.
Commit Boundary (issue #846)
spec-kitty agent mission create no longer auto-commits spec.md. The empty
template is written to disk untracked at create time; you are responsible
for committing it after writing substantive content.
"Substantive content" for spec.md means at least one Functional
Requirements row (FR-###) whose description is real (not a template
placeholder like [NEEDS CLARIFICATION …], [e.g., …], or a bare user-story
scaffold). Section presence is the only signal — adding 300 bytes of arbitrary
prose without an FR row does not count as substantive.
Workflow:
- Run
spec-kitty agent mission create …. Note thatspec.mdis left untracked. - Populate
spec.mdwith real Functional / Non-Functional / Constraint rows. - Commit
spec.mdyourself using the mission-aware entrypoint. To capture post-create fields such aspr_bound, pending-origin binding, or an optionalsource_description, always commitspec.mdandmeta.jsontogether:
Planning/spec artifacts stay in the primary partition and never transit the coordination worktree. On the dedicated non-protected topic (PR) branch this commit is direct. If routing reports a protected-ref refusal, stop and repair branch placement; do not retry against the protected ref or claim that the coordination branch is a fallback.spec-kitty spec-commit --mission <slug> --message "Add spec for mission <slug>" \ <mission_dir>/spec.md <mission_dir>/meta.json - Only then will
spec-kitty agent mission setup-planaccept the spec phase as complete; otherwise it returnsphase_complete=falsewith ablocked_reasonmentioning "committed AND substantive".
DO NOT
- Do not mix functional, non-functional, and constraint requirements in one list.
- Do not emit requirements without stable IDs (
FR-###,NFR-###,C-###,SC-###, with an optional single lowercase letter suffix such asFR-###a). Cite another mission's ID as<mission-slug>#<ID>;setup-planrefuses a malformed declared ID withSPEC_REQUIREMENT_IDS_INVALID. - Do not leave requirement status fields empty.
- Do not write non-functional requirements without measurable thresholds.
- Do not proceed to planning with unresolved requirement quality checklist failures.
Issue-Matrix Approval Heads-Up (non-gating, #3469)
If the spec cites a GitHub issue number (#NNNN), a bare/unmarked reference will later
require an issue-matrix row before its owning work package can be approved. A
context-only citation (e.g. Follow-up:, see #, parent, epic) or a PR/commit
reference (PR #NNNN, a /pull/NNNN URL) is non-gating and needs no row; if an issue
genuinely owes the mission no work, it can later be recorded with the not-applicable
verdict. This is informational only — it does not gate /spec-kitty.specify.
Charter Context Bootstrap (required)
Before discovery questions, load charter context for this action:
spec-kitty charter context --action specify --json
- If JSON
modeisbootstrap, treat JSONtextas the initial governance context and consult referenced docs as needed. - If JSON
modeiscompact, proceed with concise governance context. - If no charter exists yet, note that and continue. Missing charter is not a
blocker for
/spec-kitty.specify.
Visual Communication (recommended)
Apply the visual doctrine when a non-trivial actor flow, domain lifecycle,
rule, or concept boundary is clearer visually. Load spk-practice-show-me and
add the smallest useful diagram. Prefer an inline Mermaid diagram; use PlantUML
only when its richer layout or DSL materially helps. Keep the visual focused on
product intent—do not introduce implementation architecture into the
specification. Requirements and acceptance scenarios remain authoritative, and
trivial content needs no diagram.
Brief Context Detection (check before discovery)
Before starting discovery, check for a pre-existing mission brief:
ls .kittify/mission-brief.md 2>/dev/null && echo "MISSION_BRIEF_FOUND"
ls .kittify/ticket-context.md 2>/dev/null && echo "TICKET_CONTEXT_FOUND"
Check in priority order:
.kittify/mission-brief.md— general plan intake (written byspec-kitty intake).kittify/ticket-context.md— tracker ticket (written bymission create --from-ticket)
If a brief file is found → Enter Brief-Intake Mode
BRIEF DETECTED: .kittify/<filename> (source: <source_file>)
- Read the full brief. Do not skim.
1b. If the brief is a structured handoff packet, adopt its IDs verbatim.
A structured packet is a Markdown file whose YAML frontmatter declares
handoff_packet: 1 (contract: docs/contracts/handoff-packet-v1.md).
Also inspect .kittify/brief-source.yaml for packet_version /
requirement_ids written by spec-kitty intake.
When a v1 packet is present:
- Use each `requirements[].id` as the `FR-###` id. Do **not** renumber.
- Copy `requirements[].statement` as the FR statement.
- Preserve `requirements[].source_id` as a trace on that FR (e.g.
`Source: TKT-1042` or an equivalent spec.md trace field).
- Adopt nested `acceptance_criteria[].id` verbatim; do not mint new AC ids
for criteria the packet already numbered.
- Adopt `constraints[].id` as `C-###` (and `NFR-###` only when the
constraint is genuinely non-functional and unnumbered).
- Treat packet quality as **Comprehensive** (0–1 gap-filling questions)
when `requirements` is non-empty.
- Unknown `handoff_packet` versions, malformed YAML, or a missing
`requirements` list are **not** packets — fall through to prose
extraction below. Do not fail specify because the overlay is absent.
Still run the one-round user confirmation in step 5. Packet intake does
not skip the discovery gate.
-
Summarise for the user. Present a single paragraph: what the brief says the goal is, who it is for, and what the key constraints are. Example: "I found a plan document from Claude Code plan mode. Here's what I understand the goal to be: [summary]. I'll extract the spec from this brief rather than running a full discovery interview." For a structured packet, name the
source_tooland how many FR ids you adopted. -
Extract requirements directly. Map the brief's content to
FR-###,NFR-###, andC-###IDs. Do not ask questions the brief already answers. Specifically extract:- Objective → Functional Requirements
- Constraints and non-goals → Non-Functional Requirements and Constraints
- Acceptance criteria → FR status and Definition of Done markers
- Risks and open questions → Assumptions or
[NEEDS CLARIFICATION: <text>] <!-- decision_id: <id> -->markers (max 3; usedecision deferbefore writing each marker)
-
Ask gap-filling questions only. Scale to brief quality:
Brief quality Discovery questions Comprehensive (objective + constraints + approach + ACs) 0–1 gap-filling questions Good (objective + constraints, no ACs) 2–3 questions Partial (goal statement only) 4–5 questions Empty / missing Proceed to normal Discovery Gate below -
Show the extracted requirement set. Present the full FR/NFR/C table to the user: "I extracted X functional requirements and Y non-functional requirements. Does this look right?" Wait for one round of confirmation. This confirmation is the discovery gate for brief-intake mode; do not write or commit
spec.mdbefore it happens unless the user explicitly asks to minimize or skip discovery. The user may correct or supplement before you write the spec. -
Write spec.md normally. Apply the same quality checklist and readiness gate as standard specify. Brief-intake mode does NOT lower the quality bar — spec.md must still pass all validation items.
-
After spec.md is committed, delete all brief files (each only if present):
rm -f .kittify/mission-brief.md rm -f .kittify/brief-source.yaml rm -f .kittify/ticket-context.md rm -f .kittify/pending-origin.yaml
What brief-intake mode does NOT do:
- Does not copy brief prose verbatim into spec.md — it extracts and structures requirements
- Does not skip the quality checklist
- Does not skip the readiness gate
- Does not require the brief to be in any particular format — Markdown prose is fine
- Does not renumber
FR-###/AC-###ids when a v1 handoff packet supplied them
If no brief file is found → Proceed with normal Discovery Gate
No change to current behaviour. Continue to the Discovery Gate section below.
Decision Moment Protocol
Before asking any interview question during this command, you MUST:
-
Run
spec-kitty agent decision opento mint a decision_id:spec-kitty agent decision open \ --mission <mission-slug> \ --flow specify \ --slot-key specify.<section>.<question-slug> \ --input-key <snake_case_key> \ --question "<question text>" \ [--options '["option1","option2","Other"]']Capture the returned
decision_idfrom the JSON output. -
Ask the question to the user in chat.
-
After the user answers, run exactly one of:
- Resolved answer:
spec-kitty agent decision resolve <decision_id> --mission <slug> --final-answer "<answer>" [--other-answer] - Deferred / skip:
spec-kitty agent decision defer <decision_id> --mission <slug> --rationale "<reason>" - Not applicable / cancel:
spec-kitty agent decision cancel <decision_id> --mission <slug> --rationale "<reason>"
- Resolved answer:
-
When deferring, write the inline marker into
spec.mdimmediately after the relevant section:[NEEDS CLARIFICATION: <brief description of what needs answering>] <!-- decision_id: <decision_id> --> -
Before declaring the interview phase complete, run:
spec-kitty agent decision verify --mission <slug>Address any findings (DEFERRED_WITHOUT_MARKER,MARKER_WITHOUT_DECISION,STALE_MARKER) before proceeding.
Important constraints:
--slot-keyformat:specify.<section>.<question-slug>(e.g.,specify.auth.strategy).--input-keyis the snake_case programmatic key (e.g.,auth_strategy).- The
decision_idon the wire is a plain ULID (26 chars). TheDM-prefix appears only in artifact filenames, not in CLI arguments. - Widening is represented by the CLI/SaaS widen flow; if that flow returns
canonical thread metadata, it must be recorded as
DecisionPointWidened. - No hosted connection is required; all operations are local-only.
Discovery Gate (mandatory)
Only after create succeeds, begin brief intake or the Discovery Gate. Before
writing substantive spec.md content, committing it, or otherwise advancing
planning, you must conduct or verify a structured discovery interview.
-
Scope proportionality (CRITICAL): FIRST, gauge the inherent complexity of the request:
- Trivial/Test Features (hello world, simple pages, proof-of-concept): Ask 1-2 questions maximum, then proceed. Examples: "a simple hello world page", "tic-tac-toe game", "basic contact form"
- Simple Features (small UI additions, minor enhancements): Ask 2-3 questions covering purpose and basic constraints
- Complex Features (new subsystems, integrations): Ask 3-5 questions covering goals, users, constraints, risks
- Platform/Critical Features (authentication, payments, infrastructure): Full discovery with 5+ questions
-
Scenario-first discovery: For any non-trivial feature, prefer concrete workflow questions over abstract opinion prompts. Ask for the primary actor, trigger, happy-path outcome, and the most common exception or branch.
-
Terminology discipline: If the request introduces business or domain terms that may drift, ask which term is canonical and which synonyms should be avoided. When relevant, carry those choices into the optional Domain Language section of the spec instead of leaving them implicit.
-
Rule probing: For workflows with approvals, validations, state changes, or compliance implications, ask what must always be true and which transitions or checks cannot be skipped.
-
User signals to reduce questioning: If the user says "just testing", "quick prototype", "skip to next phase", "stop asking questions" - recognize this as a signal to minimize discovery and proceed with reasonable defaults.
-
First response rule:
- For TRIVIAL features (hello world, simple test): Ask ONE clarifying question, then if the answer confirms it's simple, proceed directly to spec generation
- For other features: Ask a single focused discovery question anchored in the primary user scenario and end with
WAITING_FOR_DISCOVERY_INPUT
-
If the user provides no initial description (empty command), stay in Interactive Interview Mode: keep probing with one question at a time.
-
Conversational cadence: After each user reply, decide if you have ENOUGH context for this mission's complexity level. For trivial features, 1-2 questions is sufficient. Only continue asking if truly necessary for the scope.
Discovery requirements (scale to complexity):
- Maintain a Discovery Questions table internally covering questions appropriate to the mission's complexity (1-2 for trivial, up to 5+ for complex). Track columns
#,Question,Why it matters, andCurrent insight. Do not render this table to the user. - For trivial features, reasonable defaults are acceptable. Only probe if truly ambiguous.
- When you have sufficient context for the mission's scope, paraphrase into an Intent Summary and confirm. For trivial features, this can be very brief. For non-trivial features, include the primary actor, trigger/success outcome, key constraint, and any explicit assumptions or deferred decisions.
- Before leaving the interview loop, do a short playback of the primary scenario, the main exception or edge case, and any rule that must always hold.
- If user explicitly asks to skip questions or says "just testing", acknowledge and proceed with minimal discovery.
Bulk-Edit Detection (mandatory check)
Before finalizing the Intent Summary, ask yourself one question:
Does fulfilling this request require changing the same existing string (identifier, path, key, label, or term) in more than one file?
Typical shapes: "rename X to Y", "the Blue feature is now the Red feature", "change the terminology from X to Y", "move package A to package B", "replace ACME with GlobalCorp everywhere in docs and UI".
If yes or uncertain: load the spk-practice-bulk-edit skill
and follow it. You will set change_mode: bulk_edit in meta.json after
mission create and produce an occurrence_map.yaml during plan. The user
does not need to know these field names — the skill teaches you the workflow.
If clearly no (a new feature with new identifiers, a bug fix that doesn't rename anything, a refactor inside one file): proceed normally.
When in doubt, treat as bulk edit. The false-positive cost is drafting one map the user approves in a pass; the false-negative cost is the silent cross-file breakage that DIRECTIVE_035 exists to prevent.
Workflow (0.11.0+)
Planning happens in the repository root checkout - NO worktree created!
- Creates
kitty-specs/<mission_slug>/spec.mddirectly in project root (the optionalNNN-prefix is display-only metadata assigned at merge time) - Commits creation metadata;
spec.mdremains untracked until it is substantive and explicitly committed - No worktree created during specify
Worktrees created later: After /spec-kitty.tasks finishes, run: spec-kitty next --agent <agent> --mission <handle>. The --mission handle can be the mission's mission_id (ULID), mid8 (first 8 chars), or mission_slug; the resolver disambiguates by mission_id and returns a structured error on ambiguity (no silent fallback). On a software-dev mission next issues the /spec-kitty.analyze step first (the required analysis gate), and only once the analysis report is current does it hand out spec-kitty agent action implement WP## --agent <name> for each WP. Each lane gets exactly one worktree, for example .worktrees/<human-slug>-<mid8>-lane-a/ (e.g. .worktrees/checkout-upsell-flow-01J6XW9K-lane-a/).
Location
- Work in: Repository root checkout (not a worktree)
- Creates:
kitty-specs/<mission_slug>/spec.md(theNNN-prefix is display-only and assigned at merge time) - Commits to: target branch (from
create --json→target_branch)
Outline
0. Establish a Provisional Mission Identity
- Before interview, derive a short provisional title and kebab-case slug from the initial request (avoid filler like "feature" or "thing"). This identity only establishes the Mission handle; do not treat it as confirmed intent.
- Resolve branch intent, then create the Mission scaffold using that provisional identity before asking any discovery or brief-intake question.
- Read the confirmed Intent Summary back to the user during discovery. It governs the substantive spec even though the Mission identity is immutable.
- Before creating, freeze the activated Mission type and the creation metadata snapshot as described above. Do not defer type selection until discovery.
The text the user typed after /spec-kitty.specify in the triggering message is the initial mission description. Capture it verbatim, but treat it only as a starting point for discovery—not the final truth. Your job is to interrogate the request, surface gaps, and co-create a complete specification with the user.
Given that mission description, do this:
- Generation Mode (arguments provided): Use the provided text as a starting point, validate it through discovery, and fill gaps with explicit questions or clearly documented assumptions (limit
[NEEDS CLARIFICATION: …] <!-- decision_id: <id> -->to at most three critical decisions the user has postponed; calldecision deferbefore writing each such marker). - Interactive Interview Mode (no arguments): Ask the single bootstrap identity prompt, create the scaffold, then use the discovery interview to elicit all necessary context and confirm it before writing substantive
spec.mdcontent.
-
Resume safely or create the Mission before interview:
-
From the initial request, derive a provisional title, purpose summary, and kebab-case slug. Resolve branch intent as required above, then call the creation command now. This must happen before the first discovery or brief-intake question so each question can open a Decision Moment against the created Mission.
-
Before
create, probe for an interrupted earlier attempt:spec-kitty agent mission check-prerequisites --mission <provisional-slug> \ --resume-probe --jsonRoute only on the structured
resume_statefield:found: reuse only whenspec_committed_and_substantiveis false and the returned Mission type and frozen creation snapshot match this invocation. The returnedtarget_branch,topology, andpr_boundmust also match the confirmed branch contract for this run. Use the returnedmission_idor exactmission_slugfor every later command. If specify is already complete, branch intent changed, or metadata differs, stop and report the existing Mission.not_found: this is the only state that authorizes a newcreate.existing: stop and report the valid merged Mission. Preserve it; never repair, remove, or reuse it as an interrupted specify scaffold.ambiguous: stop and ask the user to select one returned candidate.malformed: stop and repair or explicitly remove the partial scaffold; do not create through it.
Any result without
resume_stateis a probe/preflight failure, not proof of absence. Stop and repair it. Never infernot_foundby parsing error prose or by observing unrelated Missions. -
If this is your first message or discovery questions remain unanswered, stay in the one-question loop, capture the user's response, update your internal table, and end with
WAITING_FOR_DISCOVERY_INPUT. Do not surface the table; keep it internal. -
Only proceed once every discovery question has an explicit answer and the user has acknowledged the Intent Summary.
-
Empty invocation rule: use the response to the one bootstrap identity prompt as the provisional slug, then create before any discovery question. Stay in interview mode until you can restate the agreed-upon description. Do not write substantive spec content while the description is missing or provisional.
-
-
Run the creation command from repo root before the interview:
spec-kitty agent mission create "<slug>" \ --mission-type "<mission-type>" \ --friendly-name "<title>" \ --purpose-tldr "<purpose_tldr>" \ --purpose-context "<purpose_context>" \ --jsonWhere
<slug>is a kebab-case version of the provisional title (e.g., "Checkout Upsell Flow" → "checkout-upsell-flow").If the user expects a pull request for this work, add
--pr-bound --branch-strategy already-confirmed. Whencurrent_is_primaryis true and they accept the recommendedfeature-branchstrategy, also add--start-branch <branch>so no mission artifacts are written on the primary branch.If this mission must preserve its lane branches and/or worktrees after
spec-kitty consolidate(e.g. for post-consolidation inspection or a PR review window), declare that up front with--retain-branchesand/or--retain-worktrees— this mints a machine-readable retention policy intometa.jsonso consolidate honors it automatically, rather than relying on a prose note nothing reads.The command returns JSON with:
result: "success" or error messagemission_id: Canonical ULID machine identity (e.g.,01J6XW9KQT7M0YB3N4R5CQZ2EX). Immutable.mission_slug: Human-readable mission slug (e.g.,checkout-upsell-flow)mission_number: Display-only numeric prefix,nullpre-merge. Assigned at merge time. Never use this as a selector or identity.mission_type: Mission type key (for examplesoftware-dev)slug: Unnumbered mission slug (e.g.,checkout-upsell-flow)friendly_name: provisional titlepurpose_tldr: provisional one-line stakeholder-facing Mission summarypurpose_context: provisional stakeholder-facing context paragraphmission_dir: Absolute path to the mission directory inside the repository root checkoutcurrent_branch: the branch you started fromtarget_branch/base_branch: deterministic branch contract for downstream commandsplanning_base_branch/merge_target_branch: explicit landing-branch aliasesbranch_strategy_summary: human-readable summary of the branch contract
Parse these values for use in subsequent steps. All file paths are absolute.
IMPORTANT: Run this command at most once for a new Mission. The JSON is provided in terminal output; preserve it to get the actual paths. Retry
createonly after a confirmed non-zero failure, and repeat the resume probe before retrying because a failed caller may still have observed a completed write. Never retry after lost, truncated, or merely unparsed success output. Immediately restate the branch contract to the user after parsing the JSON:- Current branch at start
- Intended planning/base branch
- Final merge target for later changes
- Whether that matches the user's intended landing branch
If the frozen
<mission-type>is notsoftware-dev, query before handoff:spec-kitty next --agent <agent> --mission <handle> --jsonThe read-only query is the replay guard:
-
If it returns
kind: "query",mission_state: "not_started", and a type-specificpreview_step(for example,researchbegins atscoping), issue that first action exactly once:spec-kitty next --agent <agent> --mission <handle> --result success --jsonFollow the returned action and
prompt_file. -
If it returns any other
mission_state,step_id,decision_id, or non-nullrun_id, a type-specific run or action already exists. Never pass--result successmerely to recover lost output: that would falsely mark the outstanding action complete. Stop this prompt and report the Mission handle plus returned runtime fields so the outstanding action can be resumed from its original host context or recovered explicitly.
Then stop executing this prompt. Do not open software-dev specify Decision Moments, write FR/NFR/C rows, create the software-dev requirements checklist, run
spec-kitty spec-commit, or runspec-kitty agent mission setup-planfor that Mission. -
Stay in the repository root checkout: No worktree is created during specify.
-
Read the files created by
create:<mission_dir>/spec.md(already created, may be empty/template-filled)<mission_dir>/meta.json(already created with mission identity metadata)
Do NOT try to read a template file. The spec structure is defined in this prompt (see sections below). The
createcommand scaffolds an initialspec.md— read it, then update it following the structure in this prompt. -
Update
<mission_dir>/meta.jsononly when needed:- Never modify identity fields from
create(mission_id,slug,mission_slug,created_at,target_branch).mission_idis the canonical ULID and is immutable.mission_numberis display-only and isnullpre-merge — do not set it by hand. - Keep
target_branchaligned to the value fromcreate --jsonoutput. Never hardcodemain. - Preserve
friendly_name,purpose_tldr,purpose_context, andmission_typeexactly as emitted bycreate; confirmed intent belongs inspec.md, not an unrecorded rewrite of lifecycle metadata. - Optionally add/update
source_description. - Ensure
vcsexists ("git"default).
Example
meta.jsonschema (identity fields that must be present explicitly):{ "mission_id": "01J6XW9KQT7M0YB3N4R5CQZ2EX", "mission_number": null, "slug": "checkout-upsell-flow", "mission_slug": "checkout-upsell-flow-01KQ7X2M", "friendly_name": "My Mission", "purpose_tldr": "Keep the mission understandable to product and executive stakeholders.", "purpose_context": "This mission exists to make the purpose of the work immediately legible to stakeholders who should not need to parse technical specification text to understand the value or expected outcome.", "mission_type": "software-dev", "target_branch": "<target-branch>", "vcs": "git", "created_at": "2026-01-01T00:00:00+00:00" }mission_numberbecomes a concrete integer only at merge time, assigned asmax(existing_numbers)+1inside the merge-state lock. Selectors disambiguate bymission_id(or its 8-char prefixmid8), never bymission_number.Do not regenerate timestamps or directory paths via shell commands.
- Never modify identity fields from
-
Generate the specification content by following this flow:
- Use the discovery answers as your authoritative source of truth (do not rely on the raw invocation text)
- For empty invocations, treat the synthesized interview summary as the canonical mission description
- Identify: actors, actions, data, constraints, motivations, success metrics
- Prefer concrete scenario walkthrough facts (actor, trigger, success outcome, exception path) over abstract restatements
- For any remaining ambiguity:
- Ask the user a focused follow-up question immediately and halt work until they answer
- Only use
[NEEDS CLARIFICATION: …]when the user explicitly defers the decision - Record any interim assumption in the Assumptions section
- Prioritize clarifications by impact: scope > outcomes > risks/security > user experience > technical details
- Fill User Scenarios & Testing section (ERROR if no clear user flow can be determined)
- If terminology precision matters, fill the optional Domain Language section with canonical terms and ambiguous synonyms to avoid
- Generate separated requirement tables: Functional (
FR-###), Non-Functional (NFR-###), and Constraints (C-###) - Label every Functional Requirement row with a trailing
Deliveryvalue and a trailingNo-op passable?mark (summary of tacticacceptance-criteria-non-vacuity:[build]new behaviour /[ratchet]pins existing behaviour /[folded]satisfied by another row;yes/no—yesmeans a do-nothing change would pass the row's check, so reword it or name a same-fixture positive control). Place both as trailing columns only — never inside or before the ID cell, never inline on an FR bullet/heading. Full label definitions and the non-vacuity rules live in tacticacceptance-criteria-non-vacuity(fetch:spec-kitty charter context --include tactic:acceptance-criteria-non-vacuity) — do not redefine them here. - Ensure each requirement entry has a status value and testable wording
- Capture rules or invariants that shape acceptance scenarios, edge cases, permissions, or lifecycle boundaries
- Define Success Criteria (measurable, technology-agnostic outcomes)
- Identify Key Entities (if data involved)
-
Update the existing
<mission_dir>/spec.mdusing the template structure, replacing placeholders with concrete details derived from the mission description while preserving section order and headings. -
Specification Quality Validation: After writing the initial spec, validate it against quality criteria:
a. Create Spec Quality Checklist: Generate a checklist file at
mission_dir/checklists/requirements.mdusing the checklist template structure with these validation items:# Specification Quality Checklist: [MISSION NAME] **Purpose**: Validate specification completeness and quality before proceeding to planning **Created**: [DATE] **Mission**: [Link to spec.md] ## Content Quality - [ ] No implementation details (languages, frameworks, APIs) - [ ] Focused on user value and business needs - [ ] Written for non-technical stakeholders - [ ] All mandatory sections completed ## Requirement Completeness - [ ] No [NEEDS CLARIFICATION] markers remain - [ ] Requirements are testable and unambiguous - [ ] Requirement types are separated (Functional / Non-Functional / Constraints) - [ ] IDs are unique across FR-###, NFR-###, C-### and SC-### entries, and match the requirement-ID grammar (optional lowercase letter suffix; `<mission-slug>#<ID>` only for another mission's ID) - [ ] All requirement rows include a non-empty Status value - [ ] Non-functional requirements include measurable thresholds - [ ] Every FR row and success criterion carries a delivery label and no-op mark - [ ] Success criteria are measurable - [ ] Success criteria are technology-agnostic (no implementation details) - [ ] All acceptance scenarios are defined - [ ] Edge cases are identified - [ ] Scope is clearly bounded - [ ] Dependencies and assumptions identified ## Mission Readiness - [ ] All functional requirements have clear acceptance criteria - [ ] User scenarios cover primary flows - [ ] Mission meets measurable outcomes defined in Success Criteria - [ ] No implementation details leak into specification ## Notes - Items marked incomplete require spec updates before `/spec-kitty.plan`b. Run Validation Check: Review the spec against each checklist item:
- For each item, determine if it passes or fails
- Document specific issues found (quote relevant spec sections)
c. Handle Validation Results:
-
If all items pass: Mark checklist complete and proceed to step 6
-
If items fail (excluding [NEEDS CLARIFICATION]):
- List the failing items and specific issues
- Update the spec to address each issue
- Re-run validation until all items pass (max 3 iterations)
- If still failing after 3 iterations, document remaining issues in checklist notes and warn user
-
If [NEEDS CLARIFICATION] markers remain:
-
Extract all [NEEDS CLARIFICATION: ...] markers from the spec
-
Re-confirm with the user whether each outstanding decision truly needs to stay unresolved. Do not assume away critical gaps.
-
For each clarification the user has explicitly deferred, present options using plain text—no tables:
Question [N]: [Topic] Context: [Quote relevant spec section] Need: [Specific question from NEEDS CLARIFICATION marker] Options: (A) [First answer — implications] · (B) [Second answer — implications] · (C) [Third answer — implications] · (D) Custom (describe your own answer) Reply with a letter or a custom answer. -
Number questions sequentially (Q1, Q2, Q3 - max 3 total)
-
Present all questions together before waiting for responses
-
Wait for user to respond with their choices for all questions (e.g., "Q1: A, Q2: Custom - [details], Q3: B")
-
Update the spec by replacing each [NEEDS CLARIFICATION] marker with the user's selected or provided answer
-
Re-run validation after all clarifications are resolved
-
d. Update Checklist: After each validation iteration, update the checklist file with current pass/fail status
-
Report completion with mission directory, spec file path, checklist results, and readiness for the next phase (
/spec-kitty.plan).
NOTE: The script creates and checks out the new branch and initializes the spec file before writing.
General Guidelines
Quick Guidelines
- Focus on WHAT users need and WHY.
- Avoid HOW to implement (no tech stack, APIs, code structure).
- Written for business stakeholders, not developers.
- DO NOT create any checklists that are embedded in the spec. That will be a separate command.
Section Requirements
- Mandatory sections: Must be completed for every mission
- Optional sections: Include only when relevant to the mission
- When a section doesn't apply, remove it entirely (don't leave as "N/A")
For AI Generation
When creating this spec from a user prompt:
- Make informed guesses: Use context, industry standards, and common patterns to fill gaps
- Document assumptions: Record reasonable defaults in the Assumptions section
- Limit clarifications: Maximum 3 [NEEDS CLARIFICATION] markers - use only for critical decisions that:
- Significantly impact mission scope or user experience
- Have multiple reasonable interpretations with different implications
- Lack any reasonable default
- Prioritize clarifications: scope > security/privacy > user experience > technical details
- Think like a tester: Every vague requirement should fail the "testable and unambiguous" checklist item
- Common areas needing clarification (only if no reasonable default exists):
- Scope and boundaries (include/exclude specific use cases)
- User types and permissions (if multiple conflicting interpretations possible)
- Security/compliance requirements (when legally/financially significant)
Examples of reasonable defaults (don't ask about these):
- Data retention: Industry-standard practices for the domain
- Performance targets: Standard web/mobile app expectations unless specified
- Error handling: User-friendly messages with appropriate fallbacks
- Authentication method: Standard session-based or OAuth2 for web apps
- Integration patterns: RESTful APIs unless specified otherwise
Success Criteria Guidelines
Success criteria must be:
- Measurable: Include specific metrics (time, percentage, count, rate)
- Technology-agnostic: No mention of frameworks, languages, databases, or tools
- User-focused: Describe outcomes from user/business perspective, not system internals
- Verifiable: Can be tested/validated without knowing implementation details
- Labelled: Append a trailing delivery-label + no-op suffix to every criterion —
— [build/ratchet/folded] · no-op passable: [yes/no]— the same label set and no-op mark used on the Functional Requirements table (defined once, in tacticacceptance-criteria-non-vacuity; see the Requirements section guidance above).
Good examples:
- "Users can complete checkout in under 3 minutes"
- "System supports 10,000 concurrent users"
- "95% of searches return results in under 1 second"
- "Task completion rate improves by 40%"
Bad examples (implementation-focused):
- "API response time is under 200ms" (too technical, use "Users see results instantly")
- "Database can handle 1000 TPS" (implementation detail, use user-facing metric)
- "React components render efficiently" (framework-specific)
- "Redis cache hit rate above 80%" (technology-specific)
04 同仓库更多 Skills
kitty-semi-auto
Use to take a tracker issue or a coherent issue-set through a governed delivery run that works autonomously but pauses to escalate the operator's decisions. Runs BOTH ways — in this session (local) or handed to a fresh cloud runner (cloud) — from one set of standing orders covering orientation, code grounding first, Mission-vs-Ops, the pause-and-escalate contract, brownfield discipline, and the canonical closeout to a draft PR the operator merges. Triggers: semi-auto, semi-auto dispatch, run this semi-automatically, deliver these issues semi-auto, hand this to a cloud runner, dispatch semi-auto.
kitty-report-debrief
Use when a maintainer asks what happened or what is open ("what landed since Friday", "WTF happened with X", a milestone overview). Collects the facts from GitHub, writes a short branded debrief for named readers, and hands the PDF back without committing it. Triggers: what landed since, WTF happened, debrief the weekend, milestone overview, exec overview of open issues, write up what shipped.
kitty-mission-from-issue
Use to take one or more GitHub issues through a full governed Mission (spec, plan, tasks, implementation, review) to a green pull request flipped to ready for review, which a maintainer then merges. Triggers: mission from issue, run a mission for this issue, drive issue to PR.
kitty-land-pr
Use when a ready-for-review pull request needs a maintainer to get it ready to merge. Rebases it, fixes and reviews it, updates its description, changelog and docs, and hands it back without merging by hand. Triggers: land this PR, landing pass, land PR.
kitty-issue-triage
Use to triage a batch of open GitHub issues on spec-kitty/spec-kitty (default, those opened in the last 7 days). Sets type, priority, domain, parent and milestone, and reports stale, duplicate or clustered issues for an operator decision. Triggers: triage issues, triage the tracker, triage new issues.
spk-team-tracker
暂无简介