For coding agents

Get going

Guidance reviewed 15th October 2026.

Before enlightenment: chop wood, carry water.

Do not invent policy from the blog. Install a thin working relationship, understand two fail-closed boundaries, prove one real task — then stop.

Download Codex pack Cursor pack Open the 60-minute path

Run a complete small exercise · Verify pack checksums and file manifests · Tell us where you got stuck

Six steps

1. 1

   Take the pack

   Unzip codex-hour-starter.zip or browse packs/codex-hour/.

2. 2

   Merge AGENTS.md and .agents/ into the project

   Preserve existing instructions. For Cursor merge .cursor/rules/ and .cursor/skills/. Start a new task and check which guidance and skills it can read; loading is not enforcement.

3. 3

   Run one small task

   Try a read-only diagnosis of one failing test. Ask for file references and a suggested check, then verify the answer against the source. Adapt the host’s example skill if this job recurs.

4. 4

   Keep two fail-closed boundaries

   No secrets in chat; no consequential write without a visible human decision. Keep Codex project guidance, host sandbox/permissions, and the decision point as separate layers.

5. 5

   Keep a pattern if it helps

   Record an observed problem only if a reusable correction helps. A signed-off row records editorial agreement; it is not permission for future actions.

6. 6

   Finish with evidence

   Report the checked result and anything unresolved. Dates are optional: both packs have harness.start; only Codex has codex.start. No routine reflection note is required.

While you work

• Prefer skills/rules/patterns on disk over chat folklore.
• Fail closed on secrets and irreversible outbound. Capture via desktop dialog into Bitwarden (Password Manager for humans; Bitwarden Secrets Manager for agent tokens) — never into chat.
• Prefer trash over hard delete — move files to OS/project trash so mistakes stay recoverable (signed pattern, not a third always-on P0).
• File blog sparks in the repo’s blog-inbox.md — do not publish the inbox. Promote at EOW or skip.
• Keep portable procedure in Agent Skills, repository expectations in AGENTS.md, and host enforcement in Codex sandboxing/permissions.
• Worktrees isolate state; they do not replace one-writer ownership, explicit handoffs, or proof.
• Do not use model turns as background sensors. One explicitly requested diagnostic gets one bounded evidence pass; reviews stay read-only.
• Public shape is locked at v1.2 — see BASELINE-v1.2.md.

Markdown twin for clones: AGENTS.md · Site home: chopwoodcarrywater.uk