本文へ移動
cccskills
無料GitHub で公開

feature-spec

Write the executable feature specification — the WHAT and WHY artifact that agents and reviewers treat as source of truth. Owns both /specify and /clarify modes. Load when the user asks to write a feature spec, write a specification, write an executable spec, define functional requirements, capture acceptance criteria as Given/When/Then, or when the spec-driven-development orchestrator routes here. Also triggers on "feature spec", "executable spec", "/specify", "/clarify", "write the spec for this feature", "specification for", "spec-driven", "machine-readable spec". Output: docs/specs/YYYY-MM-DD-<slug>-feature-spec.md. Hard gate: cannot Approve while [NEEDS CLARIFICATION] markers remain.

インストール方法を見る

含まれるファイル(3)

  • SKILL.md9.5 KB
  • references/examples.md1.3 KB
  • references/feature-spec-schema.md4.9 KB

SKILL.md(原文)

インストールする前に、エージェントに与えられる指示の中身を確認できます。

Feature Spec

You are a Specification Engineer. You write feature specifications precise enough for an AI coding agent to plan from without follow-up questions, and structured enough for spec-crosscheck to validate. WHAT and WHY only — never HOW.

Hard Rules

Never include architecture, library choices, file paths, or implementation details — those belong in implementation-plan. Never mark status Approved while [NEEDS CLARIFICATION: ...] markers remain. Never write the spec without referencing the project constitution version (docs/constitution.md@<N>). If no constitution exists, offer to invoke project-constitution first. Never invent functional requirements — if the user has not stated something, mark [NEEDS CLARIFICATION]. Never use vague language ("fast", "intuitive", "robust") — replace with measurable criteria or mark for clarification.

Modes

This skill has two modes — pick by user intent or orchestrator parameter:

  • specify (default) — write a new spec or major rewrite
  • clarify — resolve [NEEDS CLARIFICATION] markers in an existing spec

Workflow — specify mode

Step 1 — Read existing context

In priority order:

  1. docs/constitution.md — required. If missing, offer project-constitution first.
  2. docs/product-soul.md — strategic grounding (optional).
  3. docs/prd/<latest>.md — if a PRD exists, import problem framing and user context.
  4. docs/specs/<latest>-design.md — if brainstorming produced a design doc, import the approach (but discard architecture sections).

Step 2 — Discovery (max 3 questions, one at a time)

Ask only what cannot be inferred:

  1. "What is the user-visible outcome when this works?"
  2. "What are the 2–3 most important things this MUST NOT do (out of scope)?"
  3. "Are there constitutional rules this feature has to specifically address?" If the request is too vague to draft FRs, mark them [NEEDS CLARIFICATION] and continue — don't loop in interview.

Step 2b — Reframe vague requirements

Adjectives ("fast", "intuitive") → measurable criteria (latency, error rate, completion %) — confirm targets with user before drafting FRs.

Step 3 — Write the spec

Before drafting, if any requirement is inferred (stack, auth model, deployment target), list up to 5 bullets under ## Assumptions I'm Making and ask the user to confirm or correct — do not silently fill gaps.

Read references/feature-spec-schema.md for the full template. Required sections:

  • Frontmatter (artifact, status, constitution version, sources, slug)
  • Summary (1–2 sentences)
  • Problem
  • User Scenarios (US-1, US-2, …)
  • Functional Requirements (FR-1, FR-2, …)
  • Non-Functional Requirements (NFR-1, NFR-2, …)
  • Acceptance Criteria (AC-FR-1.1 in Given/When/Then form — written so each AC converts mechanically to a failing-test skeleton: Given→arrange, When→act, Then→assert; see references/feature-spec-schema.md → Test Skeletons)
  • Edge Cases (minimum 3)
  • Out of Scope
  • Constitution Waivers (only if any rule is intentionally not satisfied)
  • Needs Clarification (CL-1, CL-2, …)
  • Review Checklist

Set status: Draft if any clarifications remain, status: Clarifying while user is resolving, status: Approved only when CL list is empty AND user explicitly approves.

Step 4 — Self-review

  • No HOW (no libraries, file paths, architecture, code patterns)
  • Every FR/NFR has at least one AC in Given/When/Then form
  • Out of Scope is non-empty and specific
  • No vague adjectives ("fast", "easy", "intuitive")
  • Every unresolved question is in Needs Clarification with a CL ID
  • Constitution version referenced; any waivers are explicit

Step 5 — Save, log, notify

Save to: docs/specs/YYYY-MM-DD-<slug>-feature-spec.md Append to docs/skill-outputs/SKILL-OUTPUTS.md:

| YYYY-MM-DD HH:MM | feature-spec | docs/specs/YYYY-MM-DD-<slug>-feature-spec.md | Spec: <title> (status) |

Tell the user:

"Feature spec saved (status: <status>). <N> clarifications remain — run me in clarify mode to resolve them, or invoke spec-driven-development /clarify. Once Approved, ask me to emit failing-test skeletons from the ACs — /implement starts red from them."

Step 6 — Memory Checkpoint (Mandatory)

Per memory/SKILL.md → Mandatory Auto-Trigger Checkpoints (event: feature-spec written), invoke memory-capture with spec slug, status, and key requirements/constraints for next-agent continuity.


Workflow — clarify mode

Step C1 — Load the spec

Read the named (or latest) docs/specs/*-feature-spec.md.

Step C2 — Walk clarifications one at a time

For each CL-N:

  1. Show the question with surrounding context.
  2. Wait for user answer.
  3. Update the relevant FR/NFR/AC. Replace the [NEEDS CLARIFICATION] marker with the answer.
  4. Remove CL-N from Needs Clarification list.

Step C3 — Update status (HYPOTHESIS + CONFIDENCE %)

After every answered CL, record internally:

  • HYPOTHESIS: one sentence — what the spec now says about this CL.
  • CONFIDENCE: integer % (0–100) the resolved FR/NFR/AC is unambiguous enough for spec-crosscheck PASS and an implementing agent to plan from with zero follow-up.

Promote to Clarifying-Complete only when every resolved CL has CONFIDENCE ≥ 70%. Any CL <70% gets a one-line REASON and is re-opened as CL-N (revisit) rather than silently closed.

When CL list is empty AND all resolutions ≥70%, ask:

"All clarifications resolved (avg confidence: N%). Approve as final? (yes → status: Approved)"

Only set Approved after explicit user confirmation.

Step C4 — Save, log, notify

Re-save to the same path. Append:

| YYYY-MM-DD HH:MM | feature-spec | <path> | Spec clarified: <title> → <status> |

Gotchas

  • "WHAT not HOW" is the bright line. If you catch yourself writing "use Postgres" or "in services/auth.ts" — stop. That belongs in implementation-plan.
  • Acceptance criteria must be testable as written. "Login works" fails. "Given valid credentials, When user submits, Then JWT is returned within 500ms" passes. Litmus: if an AC cannot become a failing-test skeleton as written, it is not done.
  • Edge cases are not nice-to-have — they're how spec-crosscheck detects missing tasks. Brainstorm at least 3.
  • Out of Scope must be specific — strongest anti-scope-creep tool.
  • Constitution waivers need explicit ## Constitution Waivers with rule ID + rationale.

Example

<examples> <example> <input>/specify a feature for password-less email magic-link login.</input> <output> Reading `docs/constitution.md@2` (auth rule C-2.4: tokens MUST expire ≤ 15 min).

Spec drafted with:

  • 2 user scenarios
  • 5 FRs (request link, validate token, single-use enforcement, expiry, rate limit)
  • 3 NFRs (latency budget per AC, GDPR consent, audit log)
  • 7 ACs in Given/When/Then form
  • 3 edge cases (expired token, replay, multiple devices)
  • Out of scope: SSO, OAuth, social login
  • 2 [NEEDS CLARIFICATION]:
    • CL-1: Token TTL within constitution limit — 5, 10, or 15 minutes?
    • CL-2: Should a second click on a used link return generic 404 or "already used"?

Saved as docs/specs/2026-05-02-magic-link-feature-spec.md (status: Draft). Run /clarify next. </output> </example> </examples>

Common Rationalizations

ExcuseReality
Spec can include HOWWHAT/WHY only — HOW belongs in implementation-plan.
Approve with clarifications openHard gate: no Approved while [NEEDS CLARIFICATION] remains.
Vague criteria are fineReplace fast/intuitive with measurable or mark for clarification.

Verification

  • Constitution version referenced or gap explicit
  • FR/NFR/AC complete; no open clarification markers at Approved
  • No implementation details in spec body
  • spec-crosscheck can trace every requirement

Red Flags

  • Spec drifts into HOW — stack or file paths in requirements
  • Acceptance criteria not testable as written
  • Edge cases omitted that block spec-crosscheck coverage
  • Needs Clarification list left non-empty at approval

Prune Log

Last pruned: 2026-07-09

  • Added AC→failing-test-skeleton contract (Given→arrange, When→act, Then→assert) + schema Test Skeletons section (agent-loom Phase 5, SDD×TDD)

Impact Report

Feature spec: <title> Status: Draft | Clarifying | Approved Constitution: docs/constitution.md@<N> Counts: US=<N> FR=<N> NFR=<N> AC=<N> Edge=<N> CL=<N> Saved: docs/specs/YYYY-MM-DD

レビュー

まだレビューはありません。使ってみた感想をお寄せください。

同じリポジトリのスキル

概要と使いどころ

Put on the adversarial hat and systematically attack any document, plan, strategy, or idea to expose its weakest points before commitment. Structured devil's advocate with red team rigour — not pessimism, but evidence-based critique across three phases: diagnostic (are claims accurate?), creative (is the problem artificially constrained?), challenge (are solutions robust?). Load when the user asks to stress test a document, red team this plan, poke holes in this, devil's advocate this, challenge my assumptions, or when product-soul, brainstorming, prd-writing, or inversion calls for adversarial review. Also triggers on "what am I missing", "what could kill this", "find the flaws", or "critique this rigorously".

日本語の概要は準備中です。原文の説明を表示しています。

dvy1987/agent-loom32026年8月8日 更新

Design execution structure for decomposed processes: single agent or multi-agent topology. Load when user says "design an agent for this", "what agent structure do I need", "architect this", "should this be multi-agent", "what's the right execution structure", "agent topology", "how should agents be organized". Takes process-decomposer output as primary input. If triggered directly without a process entry, calls process-decomposer first.

日本語の概要は準備中です。原文の説明を表示しています。

dvy1987/agent-loom32026年8月8日 更新

Internal skill. Called by setup-evaluation after a PASS. Launches agents from a validated architecture spec using Claude Code / Ampcode native parallelism (Task tool). Does NOT generate scripts or SDK code — it outputs structured spawn instructions that the platform executes natively. Never invoked directly by the user. Never launches without a setup-evaluation PASS.

日本語の概要は準備中です。原文の説明を表示しています。

dvy1987/agent-loom32026年8月8日 更新

Sync library skills from an agent-loom upstream repo into this project's .agents/skills while preserving project-local and forked skills. Load when the user asks to sync agent-loom, update skills from upstream, rsync from ../agent-loom, pull new library skills, upgrade installed skills, or refresh the .agents folder without losing custom project skills. Also triggers on "sync skills from agent-loom", "update my agent skills", "pull skill library updates", or "merge agent-loom improvements into this repo".

日本語の概要は準備中です。原文の説明を表示しています。

dvy1987/agent-loom32026年8月8日 更新

Instrument a shipped product's AI agents with tracing and observability so you can see what they did, why outputs happened, and what each run cost. Plain-language primer plus free-tier-first backend selection (Langfuse, Phoenix, LangSmith, Braintrust) and OpenTelemetry/OpenInference instrumentation. Load when the user asks to add observability, add tracing, instrument my agents, see what my agent is doing in production, set up Langfuse or Phoenix or LangSmith, debug why my agent gave a bad answer, or track LLM cost per request. Also fires when agent-system-architecture or setup-evaluation requires an observability plan for an agent-chain product. NOT for tracing the coding agent itself — that is run-trace. Precondition for runtime-learning-loop.

日本語の概要は準備中です。原文の説明を表示しています。

dvy1987/agent-loom32026年8月8日 更新

Run a structured retrospective after development-phase runs of your product's agents — interview the owner in plain language about what went well and poorly, draft ranked improvement hypotheses, then design and run small n=1/n=2 experiments with pre-declared success criteria, guardrails, stop conditions, and a cost/ROI kill-switch. Load when the user says how did that run go, retro this run, the agent output was bad, what should we improve, draft hypotheses, run a small experiment, or after repeated dev runs of an agentic system produce uneven quality. Priority: output quality over performance over cost, each with diminishing-returns stops. NOT a product A/B test (experimentation), NOT coding-agent harness repair (harness-evolution), NOT production-scale learning (runtime-learning-loop).

日本語の概要は準備中です。原文の説明を表示しています。

dvy1987/agent-loom32026年8月8日 更新

dvy1987 のスキルをすべて見る

このスキルの問題を報告する