Write, compress, audit LLM inputs—system prompts, AGENTS.md, SKILL.md. Load before editing any text a model reads, even where the change is mostly code. Use also when an agent ignores instructions, over-searches, stops early, or drifts scope.
Verify changes made to JRI source. Use when making non-trivial logic changes, such as new features, bug fixes, and refactors. Skip for documentation or copy changes.
Work until your changes are in production. Use when a change is written and must reach users—pull request, verification, merge, release.
Grill the user relentlessly about a plan, decision, or TODO until the design is agreed, then build and ship it on their word. Use to stress-test thinking, on any 'grill' trigger phrase, and before building anything the user has not decided.