Every later step in this guide is independently adoptable, and the whole guide assumes you’d rather understand a pattern than install it blind. This step is for the opposite mood: one paste that sets up the baseline kit, confirming each file with you before it’s written.
What it sets up (each explained properly in its own step):
- A working agreements section in your
CLAUDE.md: one-line gates distilled from failures that actually happen (step 1). - The two core verification gates: baseline before fix, small proof before long cycle (step 2).
- The two failure lists installed as review passes you can invoke by name (the prose list, the workflow list).
The paste
Set up my coding-agent baseline. Work one step at a time, show me every
file before writing it, and wait for my ok between steps.
1. Create (or extend) this project's CLAUDE.md with a section
"## Working agreements" holding five one-line gates:
- Baseline before fix: reproduce the bug and show the failing number
before changing code. A probe that comes back clean on its first
run is broken until proven otherwise.
- Small proof before long cycle: before any full build, render, or
deploy, run the smallest dry run that could falsify the change.
- Serve, don't describe: if the deliverable can be a running page,
render, or URL, deliver that, not prose saying it works.
- No billed calls in tests: stub every metered client; grep for live
client construction before the first test run.
- Name forks: when you hit a genuine design fork, stop and present
the options instead of picking silently.
2. Interview me, one question at a time, about failures that have
actually bitten me beyond these, and add at most three more gates in
the same one-line style. Skip anything that hasn't happened to me.
3. If I'm running Claude Code and want the failure lists as invocable
review passes: clone https://github.com/chronick/lemon-agent and
copy skills/prose-failure-list and
skills/agent-workflow-failure-list into ~/.claude/skills/. If the
repo isn't reachable, skip this and say so plainly.If you’d rather go manual
Read step 1 and step 2 and make the two edits yourself: about ten minutes, and you’ll know exactly what your agent is now promising.