Trim your AI instructions to unlock better model performance

Executive overview

As AI models improve, old instructions written to correct earlier mistakes become liabilities. Rules added incrementally to fix one-off errors accumulate into bloated prompts that constrain smarter models.

The fix is a regular instruction detox: a structured review that cuts stale, contradictory, and redundant rules. Fewer, sharper rules let newer models perform at their actual capability.

Deleting instructions is often the fastest way to improve AI output quality.

The three types of instruction rot

• Stale instructions — rules that matched an old process but were never updated when the process changed
• Contradictory instructions — rules that can't be followed simultaneously (e.g. "be concise" alongside "be thorough")
• Redundant instructions — rules that modern models follow correctly by default, making the rule unnecessary or restrictive

The quarterly (or monthly) detox: five steps

1. Pick 3–5 use cases — choose by frequency (daily/weekly tasks), risk (client-facing, financial), or leverage (high revenue impact or time savings)
2. Read the prompt line by line — don't skim; ask three questions per rule:
   • Does this still match how I actually work today? (catches stale rules)
   • Can I follow all these rules at the same time? (catches contradictions)
   • Would AI do this wrong if I deleted this rule? (catches redundancy)
3. Have AI review the prompt — paste the prompt with a review request that specifies current model names and today's date; ask it to flag contradictions and redundancies, quote the specific instructions at issue, and provide a revised version with strikethroughs for deletions and bold for additions
4. Run the deletion test — remove rules flagged as redundant or contradictory one by one; test output after each removal; keep a rule only if its removal degrades quality; expect 30–50% of rules to be safely removable
5. Progressive disclosure (optional) — instead of loading all instructions at once, tell the AI which knowledge files or subfolders to consult for specific tasks; use skills that expose only a title and description by default, with full detail retrieved on demand

Two questions before adding any new rule

• Did the AI actually make a mistake, or are you adding the rule just in case? If no mistake occurred, don't add the rule.
• Can you sharpen an existing rule instead of adding a new one? Edit rule 3 rather than adding rule 26.