skill-degradation-checklisted

# Take the data away and see what it says The failure mode that matters most in a skill is not being wrong. It is being confidently complete when it should have been partial. A skill wired to an analytics API, run with no API connected, has two options. Say "no data source connected, here is what I could not check". Or produce the report anyway with plausible numbers in the gaps. The second is indistinguishable from the first to anyone reading the output, which is exactly what makes it dangerous. ## The refusal **A skill that fabricates on a missing input fails, regardless of how good its output is when everything is connected.** There is no partial credit here. ## Procedure Run the skill four times, removing something each time. 1. **Everything connected.** Baseline. Keep the output. 2. **Data source removed.** Disconnect the API, remove the credentials, or point it at an empty dataset. Does the output name the gap, or fill it? 3. **Input file missing or empty.** Give it nothing to work from. A skill that produces a full analysis of an empty file is producing it from the model's priors. 4. **Input present but out of scope.** Feed it something it was not built for. Does it decline, or does it stretch? ## What passing looks like - the missing input is named explicitly, in the output, not in a log - affected sections are marked incomplete rather than omitted, so the reader can see what is missing rather than not noticing it is gone - no number appears tha