Interactive walkthrough that suggests improvements to make a project more efficient to work on with AI. Each suggestion is optional — present them one at a time, explain the benefit, and let the user accept, skip, or customize.

How to Invoke

/project-setup              # walk through all suggestions
/project-setup status       # show which suggestions are already adopted

Process

Before starting, read the project's existing CLAUDE.md, README.md, CONTRIBUTING.md, and any config files to understand what's already in place. Don't suggest things the project already does.

Present all applicable suggestions at once as a numbered list with a one-line description each. Default all to yes. Tell the user to reply with the numbers they want to skip, or just confirm to accept all. Then implement everything that wasn't skipped in one pass.

This is faster than going one-by-one — the user can scan the full list and opt out of what they don't want rather than answering ten separate questions.

Suggestions

Work through these in whatever order makes sense for the project. Skip any that are already in place.

1. Bug Tracker / Todo List

Pitch: A single place to track bugs, tasks, and feature requests so nothing falls through the cracks. AI agents can read and update it autonomously.

Implementation:

If the project already uses GitHub Issues, Linear, Jira, or similar — use that. Note the convention in CLAUDE.md.
Otherwise, create a TODO.md in the repo root with this format:

# Todo

<!-- Format: [status] P<priority> (category) Title -->
<!-- Status: [ ] open, [~] in progress, [x] done, [-] won't fix -->
<!-- Priority: P0 critical, P1 high, P2 medium, P3 low -->
<!-- Category: bug, feature, chore, docs -->

## Open

- [ ] **P2** (bug) Sidebar flickers on window resize
  Fix: Debounce the resize handler in `src/layout.ts`

- [ ] **P1** (feature) Add keyboard shortcuts for navigation
  See discussion in #42

## Done

- [x] **P1** (bug) Login fails silently on expired token — 2026-03-28
  Resolved: Added token expiry check in `auth.ts:validateSession()`, shows error toast

Key points:

Visually scannable — status/priority/category on one line, details indented below
"Done" entries include date and brief resolution note (what was done, not just "fixed")
Keep it flat and simple — no YAML frontmatter, no complex metadata
Compatible with the /todo and /bug-bash skills

Add to CLAUDE.md:

## Bug tracking
Bugs and tasks are tracked in `TODO.md`. Use `/todo` to add entries and `/bug-bash` to work through them.

2. Engineering Diary

Pitch: A narrative log of how the project evolved — the why behind changes that isn't obvious from code or commit messages. Useful context for AI agents resuming work and for humans onboarding.

Implementation: Create DIARY.md in the repo root:

# Engineering Diary

Latest entries first. Record significant decisions, architecture changes, and non-obvious context.

---

## 2026-03-31 — Initial project setup

Set up the project with [brief description]. Key decisions:
- Chose X over Y because...
- Structure follows...

Add to CLAUDE.md:

## Engineering diary
Maintain `DIARY.md` — add an entry when making significant changes, architectural decisions, or non-obvious tradeoffs. Latest entries at top. Write in narrative form, not bullet dumps. Focus on *why* and *context*, not *what* (that's in the commits).

3. Changelog

Pitch: A simple, human-readable log of what changed and when. Unlike git log, it's curated — only meaningful changes, no merge commits or fixups. Easy for anyone to scan without touching git.

Implementation: Create CHANGELOG.md in the repo root:

# Changelog

## 2026-03-31

- Add keyboard shortcuts for navigation
- Fix sidebar flicker on window resize

## 2026-03-28

- Initial release

Key points:

Grouped by date (newest first), one bullet per change
Short descriptions — what happened, not how
No commit hashes, no authors, no version numbers unless the project does releases
Update it with every commit

Add to CLAUDE.md:

## Changelog
Update `CHANGELOG.md` with every commit. Format: grouped by date (newest first), one bullet per change with a short description. Keep it human-readable — no commit hashes, no authors.

4. Regular Scorecard

Pitch: Periodic code quality audits catch problems before they accumulate. The /scorecard skill grades your codebase across 13 dimensions.

Add to CLAUDE.md:

## Code quality
Run `/scorecard` periodically — after completing a feature, before major PRs, or when onboarding to assess health. Address critical findings before moving on.

5. Atomic Commits

Pitch: Small, focused commits are easier to review, revert, and understand. AI agents naturally batch work — this rule keeps commits clean.

Add to CLAUDE.md:

## Commits
Break work into small atomic commits — one logical change per commit. Don't bundle unrelated changes. A bug fix, a new feature, and a refactor are three commits, not one.

6. Test and Lint Before Committing

Pitch: Catch breakage before it enters the history. Simple rule, big payoff.

Implementation: First, identify the project's test and lint commands (look at package.json scripts, Makefile, CI config, etc.). Then add to CLAUDE.md:

## Pre-commit checks
Always run tests and linting before committing:
\`\`\`bash
<test command>    # e.g. npm test, pytest, go test ./...
<lint command>    # e.g. npm run lint, ruff check, golangci-lint run
\`\`\`
Do not commit if tests fail or lint errors are present. Fix first.

Adapt the commands to whatever the project actually uses. If there's no test/lint setup, flag that as a gap and offer to help set it up.

7. Test-First Development

Pitch: Writing a failing test before implementing forces clear thinking about expected behavior and gives you a definitive "done" signal. Balance fast unit tests with thorough end-to-end tests.

Add to CLAUDE.md:

## Test-first
Before implementing a feature or fix:
1. Write a test that captures the expected behavior
2. Run it — verify it **fails** (if it passes, the test isn't testing the right thing)
3. Implement until the test passes
4. Keep a healthy mix: fast unit tests for logic, end-to-end integration tests to validate it actually works in context

Don't skip step 2 — a test that never failed never caught anything.

8. Keep README Current

Pitch: Stale docs are worse than no docs — they mislead. Updating docs at commit time costs 30 seconds; fixing confused users costs hours.

Add to CLAUDE.md:

## Documentation
Update README.md (and any relevant docs) before committing if the change affects:
- Public API, CLI interface, or configuration
- Setup/installation steps
- Feature behavior visible to users

9. Encode Preferences into Rules

Pitch: When you correct the AI or express a preference during a session, capture it permanently so it doesn't need to be repeated.

Add to CLAUDE.md:

## Evolving preferences
When the user expresses a coding preference, convention, or correction during a session, offer to encode it into this CLAUDE.md file so it persists across sessions. Examples: naming conventions, preferred libraries, architecture patterns, things to avoid.

10. Mistake Retrospectives

Pitch: When the AI makes a mistake — especially "I forgot to do X" — treat it as a process problem, not a one-off. A quick retrospective and a rule change prevents recurrence.

Add to CLAUDE.md:

## Mistake retrospectives
When you make a mistake (especially forgetting something the user asked for):
1. Acknowledge it directly
2. Identify the root cause — why did this happen? (e.g. no checklist, unclear convention, missing rule)
3. Suggest a concrete project change to prevent recurrence (add a rule to CLAUDE.md, add a pre-commit check, create a checklist in the relevant skill)
Don't just apologize — fix the system.

10. Multiple Request Organization

Pitch: Allow user to batch up ideas and braindump quickly, but Claude to work through in an organized and diligent manner.

Add to CLAUDE.md:

## Completing requests

When the user gives multiple requests:
1. Queue them mentally but complete ONE fully before starting the next
2. "Complete" means: code written, built, deployed to Docker, tested with rodney, verified working
3. If a request involves UI: screenshot the result and confirm it matches what was asked
4. Never mark something done until you've verified it works end-to-end
5. If you can't complete a request in one go (e.g., blocked by data issues), say so explicitly rather than half-doing it and moving on
6. If multiple requests conflict or depend on each other, state the dependency and ask which to prioritize

Anti-patterns to avoid:
- Starting 4 things, finishing 0
- Saying "let me commit this and move on to..." when the current thing isn't verified
- Adding code that changes behavior without testing the behavior changed
- Reacting to new user messages mid-task instead of finishing current work first
- Saying "let me ignore X for now" — either fix it or explicitly tell the user it's queued

Status Check

When invoked with status, read the project's CLAUDE.md and check which of the above are already in place. Present a simple checklist:

Project setup status:

  [x] Bug tracker — using TODO.md
  [x] Engineering diary — DIARY.md exists
  [ ] Changelog — no CHANGELOG.md
  [ ] Regular scorecard — not mentioned in CLAUDE.md
  [x] Atomic commits — rule in CLAUDE.md
  [x] Test before commit — rule in CLAUDE.md
  [ ] Test-first development — not mentioned
  [x] Keep README current — rule in CLAUDE.md
  [ ] Encode preferences — not mentioned
  [x] Mistake retrospectives — rule in CLAUDE.md
  [ ] Multiple Request Organization - not mentioned

6/10 adopted. Want to add any of the missing ones?

Guidelines

Don't push. Present each suggestion neutrally. If the user says no, move on without arguing.
Adapt to the project. A tiny script doesn't need a diary. A mature project probably has most of this. Scale suggestions to the project's stage and complexity.
Implement immediately. When the user accepts a suggestion, do it right then — create the file, update CLAUDE.md, etc. Don't defer.
Detect existing conventions. If the project already has a CHANGELOG.md, don't suggest DIARY.md as something separate — suggest adapting what exists.
Keep CLAUDE.md clean. When adding rules, integrate them with existing sections rather than appending disconnected blocks. Maintain a coherent document.