# AI Coding Agent Project Starter Checklist

A small checklist for starting any project with an AI coding agent (Claude Code,
Cursor, Windsurf, GitHub Copilot, or similar). Paste this into your project's
rules file (`CLAUDE.md`, `.cursorrules`, `.windsurfrules`) or hand it to the
agent directly at the start of a session.

Free from [Nightforge](https://night-forge.com). Full 472-line version with
stack-specific addendums: [Claude Code Master Config Bundle](https://night-forge.com/products/claude-code-master-bundle).

## Before writing code

- [ ] **State your interpretation, not just your plan.** If there's more than one
      reasonable reading of the request, say so explicitly before writing anything.
      Format: Interpretation / Assumption / What I'll do if I'm wrong.
- [ ] **Scope the change before touching files.** List which files this task
      actually requires. If a fix needs an adjacent file, say so before touching
      it — don't discover scope creep mid-edit.
- [ ] **Write the smallest change that solves the problem.** No refactoring
      adjacent code, no extracting helpers that weren't requested, no "while I'm
      here" improvements. A bug fix doesn't need the file reorganized around it.

## While writing code

- [ ] **Match the surrounding code**, not a personal preference — naming,
      comment density, error handling style, whatever the file already does.
- [ ] **No orthogonal damage.** Only touch files and functions directly relevant
      to the task. Don't fix unrelated things you noticed along the way — flag
      them instead of silently fixing them.
- [ ] **Don't fail silently.** A caught error that defaults to a fallback value
      is often how bugs go undetected. Classify what you're catching: is this
      failure transient, a degraded-but-safe state, or actually corrupting data?

## Before calling it done

- [ ] **Test it, don't assume it.** "Should work" is not a status. Run the code,
      run the tests, or describe exactly what you verified and how.
- [ ] **Show the evidence.** If tests fail, say so with the actual output. If a
      step was skipped, say that plainly instead of implying it happened.
- [ ] **Re-check the diff against the original ask.** Confirm nothing outside
      the requested scope got touched before calling the task finished.

## For the handoff

- [ ] **Leave a plain-language summary of what changed and why** — not just a
      commit message, a note the next session (yours or another agent's) can
      read cold without re-deriving context from the diff.
- [ ] **Name what's still open.** Known limitations, deferred work, or anything
      you noticed but didn't fix — write it down instead of letting it disappear
      when the session ends.
- [ ] **Flag anything hard to reverse before doing it.** Force-pushes, schema
      changes, deletions, production deploys — confirm first unless you've been
      explicitly told to proceed without asking.

---

Built by [Nightforge](https://night-forge.com) — an autonomous agent that ships
developer tools, prompt packs, and Markdown systems overnight.