Adr generator

ADR Generator

Generates and maintains Architecture Decision Records -- one-per-decision Markdown docs answering the question your future self will ask: "Why the hell did I do it this way?"

Uses the Michael Nygard format (industry standard). Each ADR lives as a numbered file in <repo>/docs/adr/. The sidecar tracks metadata only; the Markdown files are the source of truth.

When to use this skill

• Recording a tech choice -- you just picked SQLite over Postgres, Next.js over Remix, Stripe over LemonSqueezy. Write it down before the reasoning evaporates.
• During architecture phase -- catalog the foundational decisions that shape the entire project.
• Mid-build pivots -- you're changing a previous decision. Supersede the old ADR so there's a paper trail.
• Pre-launch documentation -- used alongside auto-docs to ensure every "why" is captured before you forget.

When NOT to use this skill

• No profile exists -- run the orchestrator first.
• Runtime configuration -- changing a feature flag or env var isn't an architecture decision.
• Bug fixes -- fixing a bug isn't a decision. If the fix involves a significant design change, that's an ADR.
• Third-party service tracking -- use integration-mapper for that. ADRs record why you chose a service, not its risk profile.

Prerequisites

• Project profile must exist: solo-dev-suite/profiles/<slug>.json
• Sidecar lives at: solo-dev-suite/profiles/<slug>.adr.json (created on first new)
• ADR files written to: <repo>/docs/adr/NNNN-<slug>.md

Methodology -- The Michael Nygard Format

Every ADR answers five questions:

1. Title -- Short, decision-framed. "Use SQLite instead of Postgres for v1", not "Database discussion".
2. Status -- proposed / accepted / superseded / deprecated. Lifecycle is forward-only.
3. Context -- What's the forcing function? What constraints or requirements drove this decision?
4. Decision -- What we're doing, in active voice. "We will use SQLite." Not "It was decided that..."
5. Consequences -- Good, bad, and neutral outcomes. Be honest about tradeoffs.

Plus two optional but encouraged sections:

• Alternatives Considered -- What else was on the table and why it lost.
• Notes -- Links, benchmarks, trade-off math, Slack threads, anything supporting the decision.

Status transitions

Allowed transitions:

• proposed -> accepted (decision confirmed)
• accepted -> deprecated (no longer applies, but not replaced)
• accepted -> superseded (replaced by a newer ADR)

Rejected transitions: anything going backwards (deprecated -> accepted) or skipping states.

Numbering

Sequential, zero-padded to 4 digits: 0001, 0002, etc. Numbers are never recycled. If ADR 3 is superseded, the next new ADR is still whatever comes after the current max.

Supersession

When ADR N supersedes ADR M:

• ADR M gets superseded_by: N and status changes to superseded
• ADR N gets supersedes: M
• Both Markdown files are re-rendered with updated status headers

Operations

1. New (create ADR)

Goal: Record a new architecture decision.

Workflow:

1. User describes the decision (or Claude drafts from conversation context).
2. Populate the five Nygard sections + alternatives + tags.
3. Persist:

   echo '<payload>' | python scripts/adr_tool.py new <slug> --from-stdin
   

Stdin shape:

{
  "title": "Use SQLite instead of Postgres for v1",
  "status": "accepted",
  "context": "Solo dev, no traffic yet, want zero-ops database.",
  "decision": "Ship with SQLite embedded. Migrate to Postgres if/when traffic justifies it.",
  "consequences": {
    "positive": ["Zero ops overhead", "No separate DB process"],
    "negative": ["Migration pain later if we outgrow it"],
    "neutral": ["SQLAlchemy abstracts most SQL differences"]
  },
  "alternatives": [
    {"name": "Postgres", "description": "Industry standard relational DB", "why_rejected": "Overkill for a solo dev with zero users"}
  ],
  "tags": ["database", "infrastructure"],
  "notes": ""
}

The tool auto-assigns the next number, derives a slug from the title, writes the Markdown file, updates the sidecar, and mirrors to the profile.

2. Show (view single ADR)

python scripts/adr_tool.py show <slug> --number <N>
python scripts/adr_tool.py show <slug> --number <N> --json

Human-readable output shows all sections. --json returns the sidecar entry.

3. List (all ADRs)

python scripts/adr_tool.py list <slug>
python scripts/adr_tool.py list <slug> --status accepted
python scripts/adr_tool.py list <slug> --tag database
python scripts/adr_tool.py list <slug> --json

Shows numbered list with status indicators. Filters by status or tag.

4. Supersede (replace an ADR)

python scripts/adr_tool.py supersede <slug> --old-number <N> --new-number <M>

Marks ADR N as superseded by ADR M. Updates both sidecar entries and re-renders both Markdown files.

5. Status (change ADR status)

python scripts/adr_tool.py status <slug> --number <N> --to <status>

Changes status with transition validation. Only allowed transitions succeed.

6. Render (regenerate all files)

python scripts/adr_tool.py render <slug> [--output-dir <path>]

Re-renders all ADR Markdown files plus index.md. Idempotent -- running twice produces identical output.

7. Delete (remove sidecar)

python scripts/adr_tool.py delete <slug> [--yes]

Removes the sidecar file. Does NOT delete the ADR Markdown files -- those are historical record and belong to the repo.

Sidecar data shape

Authoritative schema: templates/adr.schema.json. Top-level keys:

• schema_version (const 1)
• project_slug (kebab-case)
• created_at / updated_at (ISO timestamps)
• adrs[] -- metadata entries
  • number, slug, title, status, created_at
  • superseded_by (number or null), supersedes (number or null)
  • tags (string array)
  • context, decision, consequences (positive/negative/neutral arrays)
  • alternatives[] (name/description/why_rejected)
  • notes

Profile mirror

After every sidecar write, profile.adr_model is updated:

{
  "total_adrs": 7,
  "accepted": 5,
  "proposed": 1,
  "superseded": 1,
  "latest_number": 7,
  "latest_title": "Switch client state management to Zustand"
}

Also updates last_skill_run["adr-generator"].

Output docs

• <repo>/docs/adr/NNNN-<slug>.md -- one per ADR
• <repo>/docs/adr/index.md -- table of all ADRs grouped by status

Files

adr-generator/
├── SKILL.md                          # this file
├── scripts/
│   └── adr_tool.py                   # new / show / list / supersede / status / render / delete
└── templates/
    └── adr.schema.json               # JSON Schema for sidecar

Testing

echo '{"title":"Use SQLite for v1","status":"accepted","context":"Solo dev, no traffic yet.","decision":"Ship with SQLite.","consequences":{"positive":["Zero ops"],"negative":["Migration later"],"neutral":[]},"alternatives":[{"name":"Postgres","description":"x","why_rejected":"Overkill"}],"tags":["database"]}' \
  | python scripts/adr_tool.py new my-project --from-stdin

echo '{"title":"Choose React over Vue","status":"accepted","context":"Need component library.","decision":"Use React.","consequences":{"positive":["Ecosystem"],"negative":["Complexity"],"neutral":[]},"alternatives":[{"name":"Vue","description":"x","why_rejected":"Smaller ecosystem"}],"tags":["frontend"]}' \
  | python scripts/adr_tool.py new my-project --from-stdin

python scripts/adr_tool.py list my-project
python scripts/adr_tool.py list my-project --status accepted
python scripts/adr_tool.py show my-project --number 1
python scripts/adr_tool.py supersede my-project --old-number 1 --new-number 2
python scripts/adr_tool.py status my-project --number 2 --to deprecated
python scripts/adr_tool.py render my-project
python scripts/adr_tool.py delete my-project --yes

Expected: new creates sidecar + ADR .md files + mirrors to profile, list shows status indicators and filters work, supersede updates both entries bidirectionally, status validates transitions, render produces index.md, delete removes sidecar but .md files survive.