Developer audits AI workflow skills, finds 25,000-token file undermines itself

A developer running nine Claude Code skills in production conducted an audit to assess their effectiveness, finding that eight performed reliably while one, called feature-loop, was significantly problematic. The feature-loop skill spanned 1,002 lines and approximately 25,000 tokens, far exceeding the 70–257 line range of the other eight skills. The audit revealed that the file was saturated with urgency markers like MANDATORY and NEVER throughout, which paradoxically stripped all rules of meaningful priority since the model had no basis to distinguish critical instructions from minor ones. Additionally, in long working sessions, the skill's instructions became buried under accumulated context tokens, causing later rules to receive less model attention than earlier ones. The findings highlight a core design principle: AI workflow instructions should be concise, with critical rules kept minimal and supplementary guidance loaded only when needed.