Update OPTIMIZE_INSTRUCTIONS with lessons from analyses 69-78

[neoneye]

Summary

Documentation-only change to OPTIMIZE_INSTRUCTIONS in identify_potential_levers.py. No prompt or behavioral changes. Captures lessons from 10 self_improve iterations (analyses 69-78):

• Field-description template lock: expanded with PR Remove template-lock anchor from review_lever #484 evidence (the fix that worked) and PR Comprehensive prompt refinement: verbatim numbers, template-lock fix, tighter targets #479 evidence (replacement phrases also lock). Documents the proven fix: strip to content-type + word count.
• Negative prohibitions: new entry documenting that "Do NOT" and "Never invent" activate the banned pattern. Cites PRs Add UUID prohibition and exact-count instruction to enrich prompts #458, Use 6-char lever ID prefix, positive framing, and exact-count instruction #460, Positive framing, consistent word counts, and stronger number-evidence constraint #475.
• Consequence parroting: new entry documenting llama3.1 call-2+ behavior.
• Field desc vs system prompt consistency: new entry documenting that haiku follows system prompt, text-completion models follow field descriptions.
• Stripping too far: new entry documenting that weak models need "one sentence" and section pointer as anchors.

No experiment needed — OPTIMIZE_INSTRUCTIONS is developer documentation only, not injected into runtime prompts.

🤖 Generated with Claude Code