菜单
Produce a decision-closed ai-sdd feature plan with a verifiable acceptance checklist, grounded in this repo's conventions and the frozen contracts.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Update the ai-sdd framework binary + skills to the latest release — an apply-on-confirm wrapper that runs `ai-sdd update --check`, surfaces the current-to-latest version transition, and (only on explicit confirmation) runs `ai-sdd update` and lands a standalone reseed commit. Agent-agnostic (claude-code or codex). Use when a teammate wants to update / upgrade ai-sdd, sees the update-available notice, or invokes `/ai-sdd-update`.
Stand up (or refresh) an ai-sdd factory for a repository so any coding agent — claude-code, codex, … — can drive it. Discovers the repo's conventions, scaffolds the .ai-sdd/ home, authors worker skills + schemas, compiles the deterministic gates, and wires provider-neutral skill surfacing (AGENTS.md + per-agent symlinks). Use when onboarding a repo to ai-sdd or re-bootstrapping after the codebase/conventions drift.
Drive a software-factory run by dispatching a sub-agent per worker — the deterministic `ai-sdd` engine plans (next), a fresh sub-agent does each worker's work via its skill, the engine gates and advances (submit), and each completed slice is committed. Agent-agnostic (claude-code or codex). Use when asked to run, drive, continue, or advance a factory run / pipeline / orchestration, or when the user mentions `ai-sdd next` / `ai-sdd submit`.
Print the diagram-driven ai-sdd workflow cheatsheet — the canonical command sequence (bootstrap → plan → run, validate/next/submit/status) that travels with the binary. Use when a user forgets which command comes next, asks "how do I drive a run / what's the workflow", or needs the command reference without leaving the repo.
Turn a complete PROGRAM brief into a runnable master plan — a decision-closed program requirements doc and a master orchestration graph whose nodes are whole sub-features (kind:pipeline) sequenced by milestone validation gates with owners. The program tier above ai-sdd-plan. Requires a real program brief (refuses a one-liner) and human approval of the program requirements draft before emitting the graph. Use when planning a multi-feature, multi-person project in a bootstrapped repo (run ai-sdd-bootstrap first); each sub-feature is then planned with ai-sdd-plan. Re-run on an existing program to amend it in place (create-vs-amend is auto-detected from disk): append sub-features/milestones and rewire pending nodes; a started node (completed or in-flight) is immutable and corrected forward with a downstream `<node>-revert` node.
Turn a complete feature brief into a runnable plan for a bootstrapped repo — a decision-closed requirements doc and an orchestration graph (slices + depends_on) the engine executes. Requires a real brief (refuses a bare one-liner — asks for one) and human approval of the requirements draft before generating slices. The planning layer between an idea and ai-sdd-run. Use when starting a new feature in a repo that already has a .ai-sdd/ (run ai-sdd-bootstrap first). Re-run on an existing feature to amend it in place (create-vs-amend is auto-detected from disk): append slices and rewire pending ones; a started slice (completed or in-flight) is immutable and corrected forward with a downstream `<slice>-revert` slice.
| name | plan-feature |
| description | Produce a decision-closed ai-sdd feature plan with a verifiable acceptance checklist, grounded in this repo's conventions and the frozen contracts. |
Read first, before planning:
MASTER_REQUIREMENTS.md — the decision-closed authoritative source (where PLAN.md/
ARCHITECTURE.md disagree, it wins)..ai-sdd/conventions/<stack>.md for this slice's vertical (stack is supplied by the slice
node: contracts | operator | coprocessor | brain)..ai-sdd/schemas/{task-telemetry,self-model,parts-catalog}.schema.yaml.Write the artifact to .ai-sdd/artifacts/feature-plan.v1.yaml. It must contain:
summary: concise description of the intended change.scope: a mapping with at least in and out lists.acceptance: one verifiable item per outcome, each { id, description } with a stable id. This
checklist is the thread — the implementer records the ids it covers in satisfies, the reviewer
returns a verdict per id.decisions: every planning decision, each status: closed. Cite the MASTER_REQUIREMENTS section
or ADR that grounds it.files: exact repo-relative files/prefixes the implementation may touch — must stay within the
slice's vertical root (contracts/ | operator/ | coprocessor/ | brain/) plus shared
surfaces (tests/, docs/, Makefile, pyproject.toml, .ai-sdd/, …).tests: the gate commands the implementer must run — at minimum uv sync, make test,
make validate, make lint (per the Verification & Reviewer Runbook).Honor the boundary rules from the convention file:
contracts — other verticals import the frozen schemas.TelemetrySource/VisionSource adapter interfaces, never a concrete
provider.parts_catalog.json + prior gap
residuals — never the oracle's hidden parameters.Do not leave open architecture questions inside the artifact. If the work needs a new architecture
decision not closed in MASTER_REQUIREMENTS.md, stop and surface the options to the maintainer
instead of guessing. make demo is an m2-milestone concern — do not add it as a per-slice gate.