skill-forge

Invoke skill-forge with the topic name, optionally adding task context. It researches the library through existing published skills and web sources, fact-checks the findings, then writes a self-contained skills/<slug>/SKILL.md returned inline for immediate use.

It triages the topic against local skills by their forge mark: full coverage returns the cached skill, partial coverage improves it in place, and third-party skills are returned as-is with an advisory. Name clashes on create abort with error F9 rather than overwriting.

Yes, through multi-skill fan-out. It detects breadth during triage, asks for task context, recommends a candidate list of focused skills, and forges each approved one via parallel worker subagents, at most five at a time.

No. It follows a synthesize-only rule: found skills serve as source material and are paraphrased, never installed verbatim. If a strong existing skill is found, it recommends it and lets the caller perform the install.

A dedicated fact-check step dispatches a reviewer over the cited findings to flag unsupported, stale, or contradictory claims. Unsupported claims are dropped or repaired, and one targeted re-research round runs if coverage is thin.

Do not use it for project-specific topics that cannot be generalized, for vague topics without a concrete subject, or to forge skill-forge itself — the recursive-forge guard refuses self-replication with error F3.