Skip to content

Releases: baron-3dl/skillc

v0.2.1

Choose a tag to compare

@github-actions github-actions released this 30 Jun 04:02

Full Changelog: v0.2.0...v0.2.1

v0.2.0

Choose a tag to compare

@github-actions github-actions released this 30 Jun 03:30

Full Changelog: v0.1.3...v0.2.0

v0.1.3

Choose a tag to compare

@github-actions github-actions released this 30 Jun 03:14

Full Changelog: v0.1.2...v0.1.3

v0.1.2

Choose a tag to compare

@github-actions github-actions released this 30 Jun 00:15

Full Changelog: v0.1.1...v0.1.2

v0.1.1

Choose a tag to compare

@github-actions github-actions released this 30 Jun 00:09

Full Changelog: v0.1.0...v0.1.1

skillc v0.1.0

Choose a tag to compare

@baron-3dl baron-3dl released this 29 Jun 23:14

skillc ships a skill as one self-building file that rebuilds itself on whoever receives it, then checks its own output against examples the author approved.

Use it

git clone https://github.com/baron-3dl/skillc ~/.claude/skills/skillc
  • Author (in Claude Code): build a self-building skill from <your skill>. It runs your working skill, you approve the outputs (the only manual step), and it writes one <name>.selfbuild.SKILL.md.
  • Share: send that one file. No service, no registry.
  • Receive: install it. On first use it rebuilds against your session, tests itself against the author's held-back examples, and reports one line: built (with a score), honest-failure (with the gap), or cannot-build (naming the missing tool).

Proven, cold (fresh sessions, blind judge)

  • voice: rebuilt and answered in voice at 0.90; under a hostile "be concise, no profanity" profile it overrode and still scored 0.90; a copy with no rebuild recipe caved at 0.04.
  • explain-plainly: built end to end by the builder, then rebuilt itself cold at 0.95.
  • commit-message: in a session with no tracker, failed loud rather than guess a ticket.
  • skillc hosts itself: packaged as a self-building file, it rebuilt its own builder and emitted valid self-building skills on Opus, Sonnet, and Haiku, and under a hostile "summarize the boilerplate" profile it overrode while stamping the recipe verbatim. The loop closed: a file it emitted then rebuilt itself and ran correctly.

The rebuild recipe seed is the one part trusted by reading, not by rebuilding (Ken Thompson's "trusting trust"): all the trust concentrates into one small readable file.