Describe the workflow. Ship the skill.

Describe

No spec writing, no prompt engineering. Hand it whatever exists — a sentence (“every Friday I clean the CRM export and email a PDF report”), a runbook URL, a compliance checklist, a screenshot next to a half-working script. The factory reads the material, uncovers the implicit requirements, and writes its own internal specification before generating anything.

Don’t even know what to build? Say one word — “freight” — or paste a week of repeated tasks. It harvests the boring, repeated chores actually worth automating, drops the ones a skill can’t ship, and shapes the rest into a spec for you.

Build and gate

Five phases — discovery, design with evals, architecture, detection, build. Delivery is blocked unless every gate passes:

• Spec validation — structure, frontmatter, naming, file references
• Pipeline check — scripts compile, dependencies declared
• Security scan — hardcoded secrets, dangerous code patterns, prompt injection in the instruction body itself (override phrases, hidden unicode, encoded blobs), and network endpoints scripts reach but never declare

Install everywhere

One skill, seventeen targets. Claude Code gets a native /plugin install path; Cursor gets .mdc rules; Copilot, Gemini, Windsurf, Cline, Codex and the rest get their own native formats from one auto-detecting installer. Your colleague types one command and has the same skill in their tool.

Evolve

Every skill is born with its own metric and keeps it for life:

• Regression gate — rollout runs the skill on golden inputs and diffs the output against promoted baselines; drift that passes every shape check still fails
• Pinned judge — subjective criteria are graded by a model pinned in the spec, and a known-bad canary must fail or the judge run is invalid
• Holdout — one golden case is never shown to any optimization loop
• Model comparison — the same suite runs once per candidate model (--rollout --model A --model B) and answers “which model should run this task, and at what price”; cost comes only from runtime-declared usage, never estimated
• Evidence, not exit codes — every failure appends its raw findings to the skill’s EVOLUTION.md; evolve.py closes detect → record → re-verify from the skill’s own root