AI Decay Audit

Models get updated. Fine-tuning shifts, safety adjustments land, default verbosity creeps up — and prompts that produced perfect output on the last version quietly start producing different output on the new one. That's not a bug; it's the nature of evolving models. But it means prompt maintenance is real work, and most people only notice the decay once something visibly breaks downstream.

This skill is a diagnostician, not a rewriter. It runs three structured tests on each prompt — baseline reproduction, instruction sensitivity, and edge-case stability — then names which of the six common decay patterns is responsible: format drift, verbosity creep, tone shift, capability changes, safety boundary shifts, or instruction sensitivity changes. The diagnosis is specific. "Your output got worse" is not a diagnosis. "Your output gained 660 tokens of hedge language because the new model defaults to more cautious framing, and your prompt has no negative instruction blocking it" is.

The fix output is rated for durability. Few-shot examples and explicit format templates are Durable — they give the model concrete targets that survive most updates. Negative instructions and length workarounds are Fragile — they fight current model defaults that may shift again. You see which is which, so you can plan re-tests accordingly. For API users, the audit also calculates the cost impact of verbosity inflation in dollars, not just tokens.

It deliberately separates decay from cross-model differences. "Works on ChatGPT but not Claude" isn't decay — it's adaptation, and the skill says so rather than running a useless audit. It also has two modes: reactive (something broke, run the full process) and proactive (nothing's broken yet, harden the prompts now). The skill asks which one you need before starting, instead of forcing the full audit on a prompt that's actually fine.