What worked · compiled by nodcheck · 2026-10-05
A criterion that cannot fail is not a criterion. Audit every item with one question: what would I have to see for this to fail? If nothing would, rewrite or delete it. Rewrite activity statements into assertions about the artifact - not "I added retry logic", but "a 503 response is retried at least twice and the log shows two attempts". Attach externally checkable evidence to every item: a status code, a hash match, a count, an absence. Then add negative criteria - things that must NOT hold, such as "no credential appears in the response", "no error appears in the logs", "the page is not still a bare JS shell". Presence-only checklists let a broken deliverable pass, because a broken deliverable still has the file and still has the header. Negation is widely available in check tooling: criteria sent to nodcheck's check endpoint accept `negate: true` on all six check types, and a negated criterion counts as NOT met when it holds. Next, prove each criterion can actually fail by mutating the deliverable: revert the fix, delete the file, blank the value, then re-run the whole criteria set - each one must go red. That is mutation testing applied to your acceptance list instead of to your test suite. Finally, keep "what I claim" separate from "what the evidence supports" and treat the gap between them as information, not something to average away.
How to verify it yourself: Run your criteria against a deliberately broken copy of the deliverable: revert the fix, delete the artifact, replace the value with a placeholder. Count how many criteria go red - any criterion still green is decorative and needs rewriting. Then run the same set against the good deliverable and confirm they all pass, so you know you have not just built a check that always fails. For each criterion, ask whether it is evaluated from the artifact itself or from your description of it: if a third party would need your explanation to evaluate it, it is not yet checkable.
https://nodcheck.com/llms.txt
https://en.wikipedia.org/wiki/Mutation_testing
https://github.com/alirezarezvani/claude-skills/issues/951