About

Skill Factory

Move a skill candidate toward a reviewable release: validate its source and evidence, package only approved files, and plan a distribution route without publishing it.

Coming soon. The current prototype supports local evidence validation, allowlisted packaging, change review, and distribution planning. It has no supported public install or purchase option and does not publish to marketplaces.

What it does

1. Pin the candidate. Tie a source revision, artifact, version, supported environments, claims, and limitations to one review. 2. Validate evidence. Check that referenced files match recorded hashes and that the reviewed artifact matches its declared inventory. 3. Package deliberately. Build an archive from an explicit file allowlist and verify the output against the inventory. 4. Plan distribution. Compare a proposed route with separately checked platform capabilities and access. The result is a local plan, not a marketplace submission. 5. Recheck changes. Read-only repository change signals can mark a candidate for review again; they do not approve or publish it.

Skills begin with real work: fulfil the user's request first, then optionally turn a sanitized, verified example into a revisable idea-pool candidate. Only selected reusable concepts become prototypes; separate transfer evaluation and ongoing dogfooding inform improvement before the existing release review. Most one-off tasks stop at the answer.

The evidence supports review and reproducibility. It does not establish that a product works for customers. A changed source revision or artifact needs fresh evidence.

Development status and requirements

The prototype is a Python command-line tool with an optional local MCP server. The core uses Python's standard library; Python 3.13 or later is required. Linux 3.13/3.14 is the current tested environment. The optional MCP server has a separate pinned dependency set.

The local distribution simulation uses a fake adapter. A ClawHub wrapper prepares or runs only a dry-run command. There is no live marketplace publisher, unattended release, or automatic proof that evidence is true.

Review limits

Product acceptance still needs a clean-user installation and an end-to-end task in every advertised environment. Review should cover authentication and data flow, network permissions, source errors and quotas, claims, dependencies, licenses, and recovery from partial failures. Unit tests and packaging checks alone do not establish customer readiness.

Keep credentials and personal source data out of release evidence. Review the exact artifact and claims before any future publishing route is enabled.