requirements（requirements 搜索）

Requirements Skill

Run a structured 5-phase requirements workflow that turns a feature request into an implementation-ready specification.

Keep SKILL.md as the workflow source of truth. Load detailed templates and canned output text from references/templates.md.

Purpose

• Clarify product and technical expectations before implementation.
• Capture assumptions explicitly.
• Produce a requirements spec that references real codebase patterns when available.

When to Use

Use this skill when:

• A user asks to clarify scope for a new feature.
• A user asks to write or refine a requirements document/spec.
• A requirements command workflow is active (requirements-start, requirements-status, requirements-current, requirements-list, requirements-remind, requirements-end).

Core Rules

1. Ask one question per message.
2. Keep each question closed-form:

- yes/no, or - bounded multiple choice (2-5 options), with a default.

1. Generate all questions for the current phase first, then ask sequentially.
2. Keep product language primary; add technical notes only when they help decision quality.
3. During requirements gathering, do not implement code.
4. Use real repository paths once discovered. Use placeholders only in reusable templates/examples.
5. Keep phase question count bounded (target 5; allow 4-7 when justified).

5-Phase Workflow

Phase 1 — Initialize Session

1. Create a requirement folder: requirements/[timestamp]-[slug]/.
2. Create initial artifacts:

- 00-initial-request.md - metadata.json

1. Update requirements/.current-requirement with the active folder.
2. Scan the codebase to identify relevant modules, data models, routes, and APIs.

If the repository structure differs, adapt paths and naming to that repository.

Phase 2 — Discovery Questions (Product Scope)

1. Draft discovery questions focused on user outcomes, scope boundaries, permissions, data impact, and external integrations.
2. Write all phase questions to 01-discovery-questions.md before asking any.
3. Ask questions sequentially, each with a proposed default.
4. Accept concise answers:

- yes/no - one option from a bounded list - unknown/default

1. Store responses in temporary working notes while the phase is in progress.
2. After all questions are asked, persist finalized responses to 02-discovery-answers.md and update metadata.json.

Phase 3 — Context Gathering (Autonomous)

1. Search code for similar features and reusable patterns.
2. Review relevant backend/frontend/data files and integration points.
3. Record findings in 03-context-findings.md:

- similar features - candidate files to modify - constraints and risks - recommended patterns

1. Perform this phase without user questioning unless critical ambiguity blocks progress.

Use whichever search/read/navigation tools are available in the current agent runtime.

Phase 4 — Detail Questions (Behavior & Design Decisions)

1. Derive follow-up questions from Phase 3 findings.
2. Write all phase questions to 04-detail-questions.md before asking.
3. Ask one at a time with defaults.
4. Keep wording PM-friendly first, then include optional technical anchor when useful.

- Good: “Should users edit this from the existing Project Settings screen?” - Optional anchor: “(Likely in apps/web/src/routes/project-settings.tsx)”

1. Store answers in temporary notes during the phase.
2. After all questions are asked, persist finalized responses to 05-detail-answers.md and update metadata.json.

Phase 5 — Generate Requirements Spec

1. Create 06-requirements-spec.md.
2. Include:

- overview - functional requirements and acceptance criteria - technical requirements - files/systems impacted - assumptions and open questions

1. Mark unresolved items as ASSUMED: with rationale.
2. Update metadata status to complete (or incomplete if explicitly paused).
3. Clear .current-requirement when the session ends.

Use spec templates from references/templates.md.

Command Behaviors

requirements-start

• Initialize the folder and metadata.
• Run Phase 1, then begin Phase 2.

requirements-status

• Show current phase, progress, recent answers, and next question.
• Resume from the next unanswered question.

requirements-current

• Show full read-only snapshot of the active requirement (all artifacts).
• Do not continue questioning unless explicitly asked.

requirements-list

• List all requirement folders with status, sorted by:

1. active, 2. complete/incomplete, 3. newest first.

requirements-remind

• Show concise rule reminders based on current phase.
• Correct process drift (for example: open-ended questions, implementation drift, missing defaults).

requirements-end

• If active requirement exists, offer:

1. generate spec now, 2. mark incomplete, 3. cancel.

• Confirm destructive actions before deleting artifacts.

Portability Rules

1. Treat project-specific paths as examples, not constants.
2. Treat companion skills as optional accelerators, not dependencies.
3. Prefer generic wording unless repository context confirms specific frameworks/tools.
4. Keep artifacts and process identical across agent runtimes.

Quality Checklist

• Question sets are complete before asking begins.
• Every asked question has a default and rationale.
• Temporary notes are consolidated into phase answer files.
• Spec clearly separates confirmed decisions vs assumptions.
• Output stays requirements-focused (no implementation).

Reference Files

• references/templates.md — canonical templates for:

- question files - answer files - context findings - metadata - requirements spec - status/list/reminder output snippets