Reporails: AI Instruction Diagnostics and Governance

what ails
claude

Reporails reads the instruction files your coding agents load (CLAUDE.md, AGENTS.md, .cursorrules) and reports where the phrasing will make the agent miss your directives. Runs locally, no model calls.

# CLAUDE.md

Please try to write good tests when possible.
Avoid using mocks if you can.

vague: agent treats this as optional

# CLAUDE.md

Write integration tests for every endpoint in `src/api/`.
*Do NOT import `unittest.mock`* — test against a real
`sqlite:///:memory:` database.

specific: agent follows the directive

CORE:C:0042 Vague instruction → see the rule

# CLAUDE.md

*Never run `black`.* The project enforces style
through pre-commit hooks.

prohibition first: the agent reads the ban before the rule

# CLAUDE.md

Lint with `ruff check --fix` in `src/` and `tests/`.
Style is enforced by pre-commit hooks. *Do NOT run `black`.*

directive first, prohibition last: read in order

CORE:D:0003 Prohibition before its directive → see the rule

# CLAUDE.md

Use reporails to check the file before committing.

bare tokens: the agent parses prose, not commands

# CLAUDE.md

Run `reporails check` on the file before committing.

backticked: the agent reads it as a command

CORE:E:0003 Should be in backticks → see the rule

# CLAUDE.md

It would generally be a good idea, where it is
reasonably practical to do so, to add tests.

18 words of hedging: the directive is buried

# CLAUDE.md

Add a test for every new function in `src/`.

one clause: the directive is unmissable

CORE:E:0004 Too long → see the rule

Run it on every commit

Wire npx @reporails/cli check --strict into a git pre-commit hook. It exits non-zero when a rule fails, so the commit is blocked until the instruction file is clean.

.git/hooks/pre-commit

#!/bin/sh
npx @reporails/cli check --strict

Exits non-zero on rule failure CI variant: npx @reporails/cli check --format github

Loop Engineering: How to Stop Your Agent Reward-Hacking Its Own Checks

The steer is the one instruction in the loop you never write down, and it is where an agent learns to game its own checks

Gábor Mészáros ·July 22, 2026 ·12 min

Reward hacking is an agent editing the test to pass instead of fixing the code. The cause nobody watches is the steer, the instruction the loop rewrites to itself between retries. How to make the steer a reduction that …

Read the article →

what ails
claude

Your agent reads a file before it writes code. How that file is phrased decides whether the agent follows your instructions or invents its own.

Every rule is checked against a corpus before it ships

Instructions decay under load. Watch it happen.

Run it on every commit

Loop Engineering: How to Stop Your Agent Reward-Hacking Its Own Checks

what ails claude

Your agent reads a file before it writes code. How that file is phrased decides whether the agent follows your instructions or invents its own.

Every rule is checked against a corpus before it ships

Instructions decay under load. Watch it happen.

Run it on every commit

Loop Engineering: How to Stop Your Agent Reward-Hacking Its Own Checks

what ails
claude