BitterPillEngineering

Run the Audit workflow, which reads every force-loaded file listed in settings.json and scores each rule against five checks: whether it echoes default model behavior, contradicts another rule, duplicates one, patches a single past incident, or is too vague to follow. Each rule gets a verdict with estimated token savings.

Use the QuickCheck workflow with a file path or pasted rule block. It runs the same five-question scoring on just that target and returns a short report listing rules to keep, cut, or sharpen with reasons.

Keep concrete, actionable, non-default rules such as check harnesses, tool contracts, output format requirements, and verified gotchas. Cut rules that restate built-in model behavior, choreograph reasoning steps, or use vague directives like personality sliders.

Do not use it to attack logical flaws in ideas or arguments — that is a red-teaming task. Also verify before deleting incident-patch rules, since some suppress repeat failures, and confirm a rule truly echoes current model defaults since built-in behavior shifts between model releases.

No. The Audit workflow produces a grouped report with verdicts and estimated savings first, and only produces cleaned copies of the files after operator approval. It also supports an advisory mechanical pre-scan via a lint tool to surface candidates cheaply.