Claude TDD Kit — Data skill for Claude Code

claude-tdd-kit

[](https://github.com/shashankreddy509/claude-tdd-kit/actions/workflows/ci.yml)

**An AI coding agent will tell you it's done when it isn't.** It writes the code, writes tests that pass because they only assert what the code already does, reviews its own diff, finds nothing, and reports success. The build is green and nothing has been proven.

This kit turns each of those claims into something that can be checked. Every stage leaves evidence on disk, and the next stage refuses to continue without it.

Before / after: shipping a ticket

	Without the kit	With tdd-pipeline
Tests	Written after the code, pass first time	Stage 1.5 runs the new tests before any implementation exists and requires a non-zero exit. Tests that already pass stop the run: "they assert existing behavior and prove nothing."
Evidence	The model says "all tests pass"	The real process exit codes go into tasks/receipts/.json: red.exit, green.exit, the command that ran, and the HEAD sha
Review	The same context that wrote the code reads it	Separate specialist reviewer agents check the diff plus its callers. Every Critical finding goes to a verifier that tries to refute it
Opening the PR	Happens whenever the model decides it's finished	/ship reads the receipt and refuses if it's missing, if stage != "complete", if sha is stale, if red.exit == 0 (the tests never failed), if green.exit != 0, if any Critical is confirmed or unverified, or if review found any 🟠 Must-fix

A model can talk its way past a rule written in markdown. It can't create a receipt file that doesn't exist or a `sha` that matches a HEAD that has moved.

---

What it is

Two Claude Code plugins, installed from one marketplace. Each one works on its own:

• dev-day is the session loop. /start-session loads the corrections y