autoresearch
Autonomous goal-directed iteration loop: modify, verify, keep/discard against a metric or a checkable success predicate, with…
公开 workflow:创建、修改或迁移项目拥有的 Agent Skill package,或对现有/外部 package 做用户明确要求、零执行的只读验证与 readiness 报告时使用;也用于按已接受的 audit finding 修复 source skill。不要用于一次性回答、解释/总结/翻译、普通代码 review、第三方 Skill 纯安装或导入、跨仓批量修改,或直接修补 generated runtime mirrors。
$ npx -y skills add leo-kuang-ai/spec-first --skill spec-write-skill --agent claude-codeHow it fires
How this skill gets triggered: by you, by Claude, or both.
/spec-write-skillContext preview
The summary Claude sees to decide when to auto-load this skill.
公开 workflow:创建、修改或迁移项目拥有的 Agent Skill package,或对现有/外部 package 做用户明确要求、零执行的只读验证与 readiness 报告时使用;也用于按已接受的 audit finding 修复 source skill。不要用于一次性回答、解释/总结/翻译、普通代码 review、第三方 Skill 纯安装或导入、跨仓批量修改,或直接修补 generated runtime mirrors。
name: spec-write-skill description: 公开 workflow:创建、修改或迁移项目拥有的 Agent Skill package,或对现有/外部 package 做用户明确要求、零执行的只读验证与 readiness 报告时使用;也用于按已接受的 audit finding 修复 source skill。不要用于一次性回答、解释/总结/翻译、普通代码 review、第三方 Skill 纯安装或导入、跨仓批量修改,或直接修补 generated runtime mirrors。
把可复用目标转成 portable、source-first 的 Skill patch,或在零写入模式下报告 package readiness;交付正确分支的结果、匹配证据与 residual risks,不把 source bytes、fixture pass 或模型自述当成语义改善。
Follows `docs/contracts/workflows/scenario-capability-matrix.md` (default). Overrides: none
先根据用户请求和已确认事实选择一个 disposition;只读取会改变该 disposition 判断的 reference,并在下列 done signal 达成后停止。
| Disposition | Entry signal and result | Required action / evidence | Done signal and failure behavior | | --- | --- | --- | --- | | Near-neighbor | 非 authoring/readiness,或只请求 audit-only quality review、纯安装、runtime mirror。已接受 finding 的 remediation 不属于本分支。输出 `base_operation=null`、`effect=not-entered`、`modifier=none`;结果为 `near-neighbor-route` 或 `refuse-generated-runtime-patch`。 | 只给 owning route / next action;安装交给 `skill-installer`,mirror 交给 `runtime-maintenance`。 | 路由后停止;不得 inventory、validator、preview 或 mutation。 | | Owner blocked | create/revise 的 owner 不唯一、跨 repo、repo-external、generated-only 或 containment 未确认。保留 `base_operation=create|revise` 与 `effect=apply`,结果为 `blocked-source-owner`。 | 读 [Authoring Method](references/authoring-method.md),给 candidate-only preview、空 would-change/command list 和唯一下一步。 | 未绑定单一 canonical source 时零 mutation;不得降格为 `not-entered` 或猜测 owner。 | | Validate-only | 用户明确检查现有/外部 package;现有 package 使用 `base_operation=revise` + `effect=validate-only`,结果为 `portable-readiness-report` 或 `trust-preflight-blocked`。 | no-follow inventory、bundled validator 与 [Delivery Gates](references/delivery-gates.md)。 | 报告后停止;不得执行目标 scripts、validator、hooks、binaries 或 lifecycle,不得跟随 symlink、读 secret-like 内容、复制、安装或写入。 | | Tier A apply | 已确认 owner 的 behavior-preserving revise,具体条件由 workbench 定义。 | 读 [Authoring Workbench](references/authoring-workbench.md),确认当前授权覆盖 exact write set、preview binding 与最窄结构验证。 | receipt/验证后 close out;承重行为变化转 full apply。 | | Full apply | 已确认 owner 的 create/revise apply,结果为 `portable-core-only`、`portable-core-with-behavior-contract` 或 `spec-first-project-profile`。 | 依次读 [Authoring Method](references/authoring-method.md)、[Authoring Workbench](references/authoring-workbench.md) 与 [Shape-Aware Evaluation Design](references/evaluation-design.md);写 core 前完成紧凑 Design Record 与最小 pre-patch eval plan。Capability Map、显式 shape/module decision 或 topology 只在它们改变 owner、consumer、resource/runtime carrier、架构或风险时展开。 | preview、授权、风险匹配验证和 source update 后 close out;缺 baseline/eval plan 时停止,缺 semantic/comparative evidence 时降级对应 claim。 |
| When this fact is needed | Read | It supports | If unavailable or inapplicable | | --- | --- | --- | --- | | 判定 workflow、owner/effect 或 portable core | [Authoring Method](references/authoring-method.md) | qualification、resolution、resource ownership | near-neighbor 路由;owner 不明时 preview-only / blocked。 | | Tier A 或 full apply | [Authoring Workbench](references/authoring-workbench.md) | Tier A,或紧凑 Design Record、按风险展开的 map/topology、preview handoff | 未完成必要 design record/preview 不得 apply。 | | prose、persona、few-shot、输出合同或 agentic loop 承重 | [Behavior Contract Design](references/behavior-contract-design.md) | 行为 delta、authority、examples、checkpoint/stop | 纯工具/schema 不读;不以 persona prose 代替行为合同。 | | Full apply 需要 baseline、protected behavior 或 eval family | [Shape-Aware Evaluation Design](references/evaluation-design.md) | 最小 pre-patch eval plan 与风险匹配的 evidence family | 未读不得开始 full apply source patch;Tier A 不触发。 | | measurable optimization 或 field feedback 改变 disposition | [Optimization And Feedback Handoff](references/optimization-and-lifecycle.md) | optimization handoff、feedback-to-regression | optimization 主意图路由 `spec-optimize`;未复现 feedback 保持 observation。 | | package/写前/closeout 验证或风险信号 | [Delivery Gates](references/delivery-gates.md) | validator、preview、risk checks、五轴 readiness | 记录 `not-run` / degraded,不伪造 gate。 | | target metadata/invocation 或本地治理/catalog/generator 改变 patch | [Target Profiles](references/target-profiles.md) 或 [Project Profiles](references/project-profiles.md) | target/project facts 和 projection | portable-only 不读 profile;owner 不明不
把 AI coding 会话变成可信、由项目拥有的变更。 spec-first 是面向 Claude Code、Codex、Kiro、Qoder、Cursor、OpenCode、ZCode 与 Pi 的仓库原生 AI Coding Harness。它把想法、需求、计划、代码、审查和知识连接成可检查的工程闭环。
Autonomous goal-directed iteration loop: modify, verify, keep/discard against a metric or a checkable success predicate, with…
Audit mobile App PRD/Figma/local-source consistency across page routes, KMP/Clean Architecture, components, analytics, i18n,…
Explore vague or ambitious ideas into a right-sized requirements-only unified plan. Use when the user wants to brainstorm, think…
Structured code review for bugs, regressions, tests, and standards. Report-only by default; apply fixes only when the current…
Internal landing helper for public workflows that already hold explicit commit and landing authorization; commits scoped changes,…
Internal commit helper for public workflows that already hold explicit commit authorization; creates scoped, value-communicating…