BitterPillEngineering

To audit over-prompting in Claude Code settings, load your instruction files to classify rules using the Five Questions framework, which identifies default behaviors, contradictions, and redundancy. The audit outputs categorized CUT, MERGE, or KEEP reports to reclaim wasted tokens.

Prompt hygiene is the practice of trimming over-prompted AI instruction sets to remove redundant or fragile rules. It improves token optimization by classifying instructions against default behaviors and contradictions, estimating token savings, and preserving high-value rules like verification harnesses.

Identify fragile scaffolding and redundant instructions by applying the Five Questions framework to your system prompt. This evaluates rules for vagueness, one-off fixes, and contradictions, generating a categorized RESOLVE, EVALUATE, or SHARPEN report for cleanup.

Yes, you can perform a fast single-file check on CLAUDE.md guidance without a full system audit. This quick check delivers a targeted keep, cut, or sharpen verdict on a specific file or rule block to immediately improve instruction quality.

When trimming over-prompted AI instruction sets, keep high-value instruction types through anti-fragile safeguards. Preserve verification harnesses, tool preferences, routing rules, and specific DO and DON'T examples to maintain model performance while cutting redundancy.

AI output degrades from over-prompting when system instructions contain redundant rules, contradictions, and vague one-off fixes. This fragile scaffolding wastes tokens and confuses the model, requiring an instruction audit to classify and cut unnecessary guidance.