Install

mkdir -p .claude/skills/verify-before-claiming && curl -sL https://raw.githubusercontent.com/LeonardLeroy/i-dont-believe-you/main/skills/verify-before-claiming/SKILL.md -o .claude/skills/verify-before-claiming/SKILL.md

Codex, Cursor, opencode and others: INSTALL.md.

What it does

Before your agent can tell you anything passed, it has to run eight shell commands against the real
diff and paste the output. If one prints a line, it is not allowed to claim success.

No dependencies, no API key, no model. One markdown file.

Pin the base before the session: export IDBY_BASE=$(git rev-parse HEAD). Without it the
commands compare against HEAD, and a single git commit by the agent moves HEAD onto its
own change, which silences all eight.

What changes

Same task, same agent. On the left, what it tells you today. On the right, what it has to
tell you once the skill is installed.

Your agent today	Your agent with the skill
Fixed the retry logic. ✅ All tests pass. Ready to merge.	Fixed the retry logic. I ran check 1 and it printed: `+it.skip('resets the backoff', ...` So I skipped a test instead of fixing it. The backoff still does not reset. Not ready.

The eight checks

#	catches
1	a test disabled with `.skip`, `@pytest.mark.skip`, `#[ignore]`, `@Disabled`…
2	an assertion that cannot fail, like `expect(true).toBe(true)`
3	more assertions removed than added
4	a test file deleted, or renamed out of the runner's glob
5	an error swallowed by an empty `catch` or `except: pass`
6	"I added tests" when no test file changed
7	a named test that existed before and does not exist now
8	a test focused with `.only`, which stops every other test in the file running

When an agent writes more code than anyone reviews, the test suite becomes the only oversight
left, and optimising for it instead of the goal is the predictable
failure. Across 86,156 agent-authored test patches,
80.2% carried weak or no assertions at all. Why these seven:
docs/why.md.

Tune it

The commands in SKILL.md are plain git and grep.
Adding a framework is one regex.

There is also an optional CLI in src/ that runs the same checks as a CI gate, for when
you would rather not depend on the agent cooperating.

Credits

The problem these checks exist for is measured in All Smoke, No Alarm: Oracle Signals in
Agent-Authored Test Code by Dipayan Banik, Kowshik Chowdhury and Shazibul Islam Shamim, and in
METR's work on reward hacking. The checks themselves are not from those papers. They are eight
things a diff can prove, cheap enough to run every turn.

License

MIT. Contributions welcome: CONTRIBUTING.md.

Star ⭐ if it saved you one merge past one "✅ All tests pass."