Skill Forge v0.4.2

Skill Forge v0.4.2 adds an optional evaluation-resource preflight and captured-access audit. Pin exact companion files, hashes, provenance and sharing constraints before comparing candidates. Observed violations fail; incomplete or absent capture remains inconclusive.

The helper validates supplied evidence. It is not a sandbox, trace collector or authorization service. Core interview behavior and prior features are preserved. Complete production access capture and downstream model-quality improvements remain unverified.

• Source: merged PR #6, commit e5631173fb2c6c91d9f851ba6a56f926cc50919d.
• Validation: 76 local tests ran (75 passed; one Windows-native skip). Independent review and reserved boundary checks passed. All six Windows/Linux/macOS and Python 3.9/3.13 jobs passed on the final PR head in the PR workflow and push workflow.
• Metadata, links, portable generation, publication selection and four clean filesystem installations passed. See evidence and limits.
• Future-session plan: supported model handoff, trustworthy access capture, native Office/cloud round trips, generated GitHub execution and model/allowance calibration. This is planned work, not active background execution.

Download the complete skill ZIP for Codex/Claude Code, the portable ZIP for another capable LLM, or the source ZIP for the repository's reviewed publication files. Verify the ZIPs using SHA256SUMS.txt. Install the complete skill folder, then open a new session and invoke $skill-forge.

Canonical skill tree SHA-256: f4cf8b9b2598b20e46ccda8a99441704172860546115c7cc0d29722a93018a3e.

Publication verification: all four public assets were downloaded and matched the prepared bytes and GitHub SHA-256 digests. All 288 source files matched the tagged checkout; all 36 skill files matched the source ZIP, skill ZIP and managed installation. The downloaded source passed release/link/portable checks and four disposable clean installations. Tag v0.4.2 and main both resolve to the source commit above. These checks cover filesystem/CLI installation; native host discovery remains separate.

Skill Forge v0.4.1 — guarded macOS REA recovery

Fixes the observed macOS SwiftBridging duplicate-definition failure for REA JavaScript analysis through an explicit guarded, process-local compiler overlay. Apple files and REA native ownership checks remain unchanged.

Verified: original JavaScript fixture plus two independently sealed fixtures passed actual REA 6.8.0 static semantic assertions with clean exits. Local suite: 67 passes, 1 existing Windows-only skip. Candidate 1 cleanup failure was rejected; the accepted candidate retains dependent files and reports uncertainty after abnormal native exits.

Core SHA-256: 850e1fd6286663bb0c7f360a2e47adca7c78822917b7ed7a2e4d8d4fad01f454. The installed personal Codex copy matches and passed the real original fixture.

Download the skill ZIP for Codex/Claude Code, portable ZIP for capable chat models, or source ZIP for development. SHA256SUMS.txt covers the three packages.

Recovery guide: https://github.com/HarryBurd33/skill-forge/blob/v0.4.1/skills/skill-forge/references/rea-macos.md
Validation: https://github.com/HarryBurd33/skill-forge/blob/v0.4.1/evals/v0.4.1/REPORT.md

Limits: specific opt-in macOS workaround, not a global Swift repair. Normal host process inspection is required; sandbox denial remains possible. Use an actual diagnostic capture. Failed/interrupted native work may leave retained temporary state. Static facts do not prove target application execution or complete call resolution. Intel Mac native recovery and other OS REA providers remain unverified. The portable text omits executable assets.

All six hosted CI jobs passed (Windows/Ubuntu Python 3.9 and 3.13, Intel macOS 3.9, ARM macOS 3.13): https://github.com/HarryBurd33/skill-forge/actions/runs/40656084712
Merged PR: #5
Release source commit: fb341a035fc1cbc50eb5d6c4e794e320cf4f1800.

Closing verification — 2026-10-28: all four public assets were downloaded again and matched their prepared bytes and GitHub SHA-256 digests. All three ZIPs matched SHA256SUMS.txt. All 284 downloaded source files matched the published checkout; skill ZIP, downloaded source, canonical source and installed personal copy shared the core hash above. The downloaded source passed release/link/portable checks and four clean temporary Codex/Claude personal/project installations. Local Git is clean at the v0.4.1 tag, with prior versions and reviewed changes preserved for rollback. Native Claude model execution remains unverified.

Skill Forge v0.4.0 — REA and Anatomy

Skill Forge now learns reusable workflows from readable source or authorized software artifacts, with a scoped REA adapter for compatible artifact-analysis tools. Relevant physical explainers automatically receive the real pinned Anatomy SVG/HTML renderer, CSS, license and self-contained model/build resources. Text/CSV workflows and explicit visual opt-outs are preserved.

Validation: all 10 fixed checks passed; six candidate model scenarios passed versus four baseline scenarios; 42 local helper tests passed with one Windows-only skip. Six hosted platform jobs cover Windows, Linux, Intel macOS and ARM macOS. Required HTML layout, keyboard and reduced-motion checks were maintainer-reported and independently reviewed; browser/version is unknown. Requested actor settings were GPT-6.1 Sol/medium; actual serving metadata was unavailable.

The laptop managed installation matches canonical core SHA-256 da8e24cf80f7dfc57597a7746ff917a22a2030adf4a7ee7983506e446380dbeb. Prior versions and evidence-gated rollback records are retained.

Limits: REA is optional and its engines/MCP registration are not installed by Forge. Real ZIP inspection succeeded; JavaScript semantic analysis remains unavailable on the tested macOS toolchain after a SwiftBridging compiler conflict. Anatomy includes the basic SVG/HTML slice rather than the full advanced toolkit. Native Claude and other-model parity remain unverified. Plain-chat prompts require supplied assets/tools for executable visual outputs.

Download the skill ZIP for Codex/Claude Code, portable ZIP for other capable chat models, or source ZIP for examples/tests/docs. Verify downloads using SHA256SUMS.txt. See README.md, docs/integrations.md and evals/v0.4.0/REPORT.md for setup, provenance, checks and limitations.

Post-publication verification: all four public downloads matched the prepared files and SHA256SUMS.txt. All 248 source-package files matched the accepted repository; the downloaded skill, canonical skill and installed copy shared the same complete core hash. The downloaded source passed clean first-use checks in four temporary Codex/Claude personal/project paths. Both merged/tag validation workflows passed; these filesystem/helper checks do not establish new native Claude or other-model execution.

Get started: README, Codex setup, Claude Code setup, validation and limits.

Skill Forge v0.3.1

A focused validation update informed by Anthropic’s Complete Guide to Building Skills for Claude.

Changes

• Optional validate --profile claude checks reserved name substrings and angle brackets in discovery metadata. Generic validation remains the default; body XML/HTML stays allowed.
• Supplied compatibility values receive 1–500-character checks.
• Original source comparison, provenance and rights limits are documented; the PDF and extracted text are not redistributed.
• Existing interview, spec-review, Skill Finder and evidence-gated update behavior is preserved.

Validation

Frozen real-CLI comparison: 26/26 cases pass, including four independently reserved cases; baseline16/26. Correctly rejected invalid metadata improves3→12. Independent review and actual automatic adoption passed. Local unit suite:30passed,1Windows-onlyskip. Source/package links and four temporary host installation paths pass.

The installed laptop skill matches canonical core SHA256 5b25ded0676dc20f8331442bfe62a4d3e17df05a9f71603cd745224c7b5a27c7.

Downloads and limits

Choose the skill ZIP for Codex/Claude Code, portable ZIP for a capable chat model, or source ZIP for development; SHA256SUMS covers all three. Setup · Source review · Validation.

These are supported-subset metadata checks, not full YAML validation, a security audit or proof of native Claude acceptance. No new interview/model-quality benchmark was run. Existing native-host limitations remain.

Verification after publication

All four public release assets were downloaded and matched local files, server digests and SHA256SUMS. All 221 selected source files match the merged source; the skill ZIP and your managed laptop installation have identical core hashes. Validation with the Claude profile, the full unit suite, portable drift checks and clean package/installation checks passed from downloaded source. Six hosted platform jobs passed on the feature commit. The canonical laptop repository is synchronized to merged commit 1dad1ed09570e151b81051d928c052a370016c88.

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.

Skill Forge v0.2.0

Skill Forge turns an idea into a reusable skill through focused questions, examples, and evidence.

What changed

• Choose the skill’s destination: local Markdown, an installable folder/ZIP, GitHub, or another implementation. Publication requires your chosen destination and authorization.
• Continue refining through dependent questions and constructive challenge while material gaps remain. Quick, thorough, expert challenge, draft now, skip, and resume remain available.
• Skill Finder delegates bounded research before implementation, inspecting useful GitHub repositories, Hugging Face resources when relevant, actual source, licenses, and fit. Limited-tool hosts get an honest direct-search or no-tools fallback.
• Save delivery, output, refinement, and research state separately. Finder ships in the installed skill and the self-contained prompt.

Choose a download

Use the skill ZIP for Codex or Claude Code, the portable ZIP for another capable chat model, or the source ZIP for development. SHA256SUMS records all three archive checksums.

Codex setup · Claude Code setup · Evaluation report

Validation: 20 helper tests; final isolated suite 12/12 cases and 50/50 conditions versus 43/50 for the prior version, with no observed regressions; six description-routing probes; clean installation/export checks. Prior failures, protocol corrections, and one unassisted source-retrieval retry are retained in the report. The passing update was adopted automatically with snapshots and rollback.

Native Claude Code execution and other model families remain unverified. Tool capabilities vary by host. Original project material is MIT licensed; third-party reuse rights remain separate.

Skill Forge v0.1.0

Skill Forge v0.1.0

First usable MIT-licensed release: create skills from ideas, improve existing skills, and evaluate improvements to the builder itself.

Includes adaptive interviews; durable private session records; a searchable question library; evidence-gated automatic adoption and rollback; Codex and Claude Code setup guides; a self-contained portable prompt; three domain examples; and public evaluation evidence.

Downloads

• skill-forge-v0.1.0.zip: installable skill folder with helpers, license and notices.
• skill-forge-portable-v0.1.0.zip: self-contained prompt for capable chat models.
• skill-forge-source-v0.1.0.zip: reviewed source, examples, tests, documentation and evidence.
• SHA256SUMS: checksums of the three ZIP files.

Executed checks and limits

18 helper tests pass. The publication checker validated 90 selected files and 93 relative links. Clean temporary installation and export checks pass. Independent model review found 14/17 behavioral cases passing versus 12/17 for the direct baseline; routing review was 6/6 in both arms. These are small smoke comparisons, not statistical superiority claims. All failures and raw synthetic outputs are preserved in the evaluation report and linked evidence.

Actual builder and target-skill candidates were automatically adopted after their checks; redundant and no-benefit candidates were rejected without changing the active version. Rollback, customization protection and private-state exclusion were executed in tests.

Native Claude Code execution and host UI discovery remain unverified. Text-only runs exposed resume-questioning, candidate-folder naming and assumption-labeling limitations. Generated skills should be validated before installation. Tool access and model behavior vary by host; no paid API or custom service is required for the core workflow.

Start with README and the Codex or Claude Code installation guide. Contributions can propose a useful question or a reproducible failure with expected behavior; private transcripts and credentials do not belong in the repository.