The original title is about coding agents and their skill descriptions being broken. Let me analyze the key facts:

Short summary

An analysis of why AI coding agent skill descriptions fail: when descriptions contain executable procedures, the model skips reading the skill body and runs a shallow version from memory. The fix is to write descriptions as pure routing triggers ('Use when...') with zero procedure verbs, include negative discriminators for neighboring skills, and test routing by feeding only descriptions to a subagent with real task prompts.

• •Skill descriptions should be routing triggers only — never contain executable procedures the model can run from memory
• •Include 'NOT for...' clauses to discriminate between neighboring skills with overlapping scope
• •Test descriptions by giving a subagent only the description list and real task prompts, then grade routing accuracy

Generated with AI, which can make mistakes.