skill-creator-trigger-eval-gotchas

Probe shadowing in parallel worker runs can make recall look artificially zero even when verbatim trigger phrases should match. Diagnosing whether the failure is caused by sibling probe collisions or the skill itself prevents unnecessary description tuning.

Wrong project root resolution in a skill evaluation loop happens when cwd-sensitive root discovery fails. You must verify the current working directory matches the expected project root before running probe-based recall tests, ensuring ground-truth validation against real skill triggers.

Probe shadowing in multi-worker skill evaluation occurs when sibling probes collide during parallel runs, causing misleading false failures. Preventing parallel-worker collisions ensures that harness results accurately reflect real recall instead of artificially zero scores from overlapping probe execution.

Validating real skill recall requires comparing harness results with ground-truth Claude invocations before adjusting descriptions. This ground-truth validation confirms whether a trigger-eval failure reflects an actual skill issue or an environment problem like Anthropic authentication errors.

Skill trigger evaluation does not work correctly with missing Python dependencies in Claude Code environments, as execution problems cause false failures. Resolving environment issues like missing dependencies and Anthropic authentication errors is required before accurate probe-based recall testing.