Refinements to LLM guardrails in `/AGENTS.md` based on real-world experience in June 2026
#3591583+ has been working well for me.
Things I've noticed during my intense use of AI in the past few weeks while working on %"Symmetric Translation: Phase 1", where I worked solely on back-end functionality:
1. `npm run lint:cspell` is almost always forgotten, probably because it currently says: _"ALWAYS run `npm run fix` when JavaScript code is updated.". (Which implies it's only necessary for front-end changes, which is not true)
2. Often, but definitely not always, are there PHPCS and/or PHPStan violations.
3. The most common PHPCS violation: 80 cols rule, and specifically for a `@todo` comment being flowed across multiple lines; even though the rule actually allows the `@todo` rule to be as long as it needs to be on a single line
4. It generates decent commits to push up to an MR, but it's very hit-or-miss (even within a single session) what the first line of a commit message looks like (the one that `git log --oneline` shows): `#<work item ID> Something` or `🤖 Something`. I noticed that the template that #3591583+ added is ambiguous.
issue