Skill Forge v0.3.0

Skill Forge turns ideas and conversations into independently usable skills, improves existing skills, and evaluates improvements to the builder.

What changed

• Shows an implementation specification and concrete verification plan before building. Ask for review first to pause; draft-now and already-authorized work can continue.
• Preserves settled decisions and corrections from conversations or project context. Skill delivery is separate from the documents, tables, or other artifacts it will produce.
• Adapts questioning depth, supports quick/thorough/expert-challenge modes, and uses examples and constructive disagreement until material gaps close.
• Uses available useful builders/reviewers with bounded tasks and recorded results. Nested Skill Finder and honest tool-free fallbacks remain included.
• Applies successful updates after predefined checks, with history, rollback and customization protection. Private state remains excluded from packages.

Downloads and setup

• skill-forge-v0.3.0.zip — installable skill folder for Codex or Claude Code.
• skill-forge-portable-v0.3.0.zip — self-contained prompt for another capable model with sufficient context.
• skill-forge-source-v0.3.0.zip — source, examples, tests, documentation and public synthetic evaluation evidence.
• SHA256SUMS — checksums for the three archives.

Codex setup · Claude Code setup · Other LLMs

Executed validation and limits

21/21 original scenarios and 84/84 required conditions passed; both descriptions routed 6/6 requests correctly. Three supplemental cases passed all 15 items. Navigation-only maintenance preserved instruction bytes, added five section links, and passed all four structural checks. Actual rollback, managed upgrade/downgrade and customization refusal passed. A fresh read-only Codex CLI 0.164.0 Sol/Medium run discovered the installed final skill and returned a plan and standalone draft.

Six hosted platform jobs passed the Windows/Linux/macOS helper, package and clean-install matrix. The final allowlist contains 210 files with 194 relative links. Native Claude, native Windows/Linux model runs and other model families remain unverified. Eight historical cached pairs use different recorded model aliases; the matching-settings robustness subset still improves. These small tests are not universal or causal superiority claims. Full report and preserved limitations.

Final installed/canonical/public skill SHA-256: 86daffd3c4e7cdeed994717466b0ba9a8e3c6c16bd596acbef3ed01114963608.

Original project material is MIT licensed; third-party reuse rights and notices remain separate. No paid API or custom service is required for the core workflow.

Verification after publication

All four assets were downloaded from the public release URLs. Their bytes match GitHub digests, local packages, and SHA256SUMS. All 210 selected source files match the release; the downloaded skill, canonical source, and personal installation share the final skill hash above. The four documented validation commands passed from the downloaded source, including four clean temporary installation paths. Release-commit CI also passed.