Design Tracker Skill

Purpose

This skill keeps the project's 要件定義書 (.claude/docs/DESIGN.md) current. DESIGN.md is the macro requirements & design document (what the project builds and why); micro work progress lives in PROGRESS.md. It covers:

• Background & purpose, scope
• Functional & non-functional requirements
• Architecture (including agent roles)
• Tech stack choices and their rationale
• Constraints, key decisions, and open questions

How This Skill Is Reached

Read this before relying on it: nothing in the repository activates this skill automatically. The description above is the entire trigger surface, and it works only through the runtime's own description-based skill selection.

• Explicit request — "record this", "add to design", "update DESIGN", "記録して", /design-tracker. This is the reliable path.
• Model invocation from the description, at the runtime's discretion. Claude Code discovers the skill through .claude/skills → .claude/skills; Codex through .codex/config.toml's path = ".claude/skills/design-tracker".
• No hook mechanism. .claude/hooks/ contains no design-tracker branch, and the words a design conversation actually uses (設計 / design / architecture) are claimed by CODEX_TRIGGERS in agent-router.py, which injects a Codex consultation nudge instead. check-codex-before-write.py also nudges when DESIGN.md is edited, but it blocks nothing — a freehand edit still lands.

So an agent in a design conversation must decide to record the decision; no automation will decide for it. If a decision was made and this skill was not reached, record it at the next checkpoint — /checkpointing describes recording decisions via update_design.py for exactly that reason.

Previous versions of this file promised proactive, automatic activation ("Do NOT wait for user to ask"). That promise was enforced by nothing, so it has been removed rather than left as a false statement in a normative document.

Workflow

Recording Decisions

1. Decide whether this is a design decision, and whether it is already recorded. Grep the target table (e.g. grep -n "^| " .claude/docs/DESIGN.md) instead of reading the whole document — the writer locates the table and heading itself.
2. Extract the decision from the conversation.
3. Map it to a section and to that section's typed input key (table below).
4. Write a per-invocation input JSON and run the shared writer (see Mechanical Update).

Sections to Update

DESIGN.md uses these fixed sections (Japanese + English headings). Every target has a typed input key, so no markdown row is ever hand-written:

Conversation Topic	Target Section	Input key	Fields
Project goals, problem, stakeholders	## 背景・目的 (Background & Purpose)	section_updates	heading, content (prose)
What is / isn't covered	## スコープ (Scope) — In / Out of Scope	section_updates	heading, content (bullets)
A feature the system must provide	## 機能要件 (Functional Requirements)	requirements	id, requirement, priority, notes
Performance, security, availability, maintainability targets	## 非機能要件 (Non-Functional Requirements)	nfr	category, requirement, metric
System structure, components, agent roles	## アーキテクチャ (Architecture) — overview + Agent Roles table	section_updates for the overview prose, agent_roles for the table	agent, role, responsibilities
Library / framework / infra choice + why	## 技術選定 (Tech Stack & Rationale)	tech_choices	area, technology, rationale, alternatives
Hard limits (technical, org, compatibility)	## 制約 (Constraints) bullets	section_updates	heading, content (bullets)
Why we chose X over Y (significant)	## Key Decisions	decisions	decision, rationale, alternatives (the date is stamped by the writer)
Things to do later, unresolved questions	## TODO / Open Questions	section_updates	heading, content (checklist)

The four prose sections have no typed key because they have no fixed shape — their content is a sentence or a bullet an agent writes, and there is no correct rendering for a script to own. Every section that does have a fixed shape (a table) has a typed key, and the writer refuses table rows passed through section_updates: it exits 2 naming the key you should have used, so the unescaped-cell and orphaned-row corruptions are now unreachable from this skill.

Choosing the section stays judgment. Rendering the row does not.

Mechanical Update

Use a per-invocation input path, never a shared one: this skill can run concurrently with other work (and inside a subagent), and two recordings sharing one input file overwrite each other. Resolve the path from the shared workspace registry rather than deriving it by hand, so the slug rule is the same one every other skill uses:

python3 .claude/skills/_shared/workspace.py \
  --skill design-tracker --title "{decision topic}" --create

That prints one JSON object whose paths.design_input is .claude/logs/design-input-{slug}.json. Use it verbatim as ${input} below. Exit 0 resolved/created · 1 bad args · 3 .claude/logs/ could not be created.

Example input (use only the keys you need):

{
  "decisions": [
    {"decision": "Use ReAct pattern", "rationale": "Better tool-use control", "alternatives": "Function calling only"}
  ],
  "tech_choices": [
    {"area": "Agent loop", "technology": "ReAct", "rationale": "Tool-use control", "alternatives": "Function calling only"}
  ],
  "section_updates": [
    {"heading": "## TODO / Open Questions", "content": "- [ ] Evaluate streaming support"}
  ]
}

Run dry-run, read the preview, then apply:

python3 .claude/skills/_shared/update_design.py --input "${input}"
# Read the file named by preview_file in the JSON output, then:
python3 .claude/skills/_shared/update_design.py --input "${input}" --apply --require-change

Completion test. "ok": true alone is not it — a duplicate or empty entry used to return ok: true with result: "no-op" and exit 0 while nothing was written. Require all of:

• result == "applied", and
• decisions_appended > 0 or some rows_appended value > 0 or sections_updated non-empty.

Always report skipped_duplicates when it is non-zero — that is the honest "already recorded" answer. --require-change makes the writer enforce the same thing: a no-op becomes ok: false and exit 2, so "recorded" can never be reported for a run that wrote nothing.

Other exit codes: 1 bad arguments or input-schema violation · 2 DESIGN.md structure invalid or missing (run /init first), a duplicate requirement ID, a table row passed through section_updates, or a no-op under --require-change · 3 DESIGN.md changed while the writer held it, or the write failed — re-read and retry.

Output Format

When recording, report concisely:

• What was recorded, and into which DESIGN.md section
• The writer's result, the appended counts, and skipped_duplicates
• Anything you decided not to record, and why

Language Rules

• Reasoning / code examples: English
• Document content: English (technical terms); Japanese descriptions are acceptable to match the existing 要件定義書 headings
• Report: follow the surrounding session's language