skillify

Run skillify on the feature. It audits against a 15-item checklist, writes the SKILL.md with frontmatter and eval contract, extracts deterministic code into scripts, runs a cross-modal eval, then adds tests and a resolver entry with real trigger phrases.

Cross-modal eval runs three frontier models from different providers (OpenAI, Anthropic, DeepSeek by default) to score a skill's output against its declared eval contract dimensions. Every dimension mean must reach 7 with no single score below 5 to pass.

Yes, skillify is idempotent. On an existing skill it detects the delta (bug fix, new input, or quality pass), edits files surgically, re-runs all tests, bumps the version, and enforces the no-regression gate against the prior eval receipt.

Skip it when the output is under 200 tokens, when the skill is a thin wrapper around a single API call, or for a pure deterministic bug fix fully locked by new tests. The Phase 7 re-verify still runs in all cases.

The no-regression gate blocks the edit from shipping. A new overall score below the prior receipt, or any dimension dropping more than 0.5, is a regression that must be re-fixed or reverted before release.

Scheduled skills can silently ship degraded output to live channels. The gate requires re-running a representative task, eval'ing the real output, and scoring at or above the previous iteration before the edit ships.