Changelog

Newest first.

Unreleased

1.0.0, 2026-10-19

• 1.0. The version now says what Skill Builder is: complete and in daily use. 18 skills, 8 agents, the engine on npm, and every tool's install and update path documented. It works the same as 0.1.1.

0.1.1, 2026-10-19

• Install covers every tool, and updating too. The README's install section adds ChatGPT and the Codex app, a terminal install of the engine from npm, Antigravity in the table of what each tool runs, and a table of how to update and remove Skill Builder in each tool.
• --help never runs a command. skill-builder release --help used to start a release. Help and --version now print and stop, whatever command comes before them.

0.1.0, 2026-10-19

• The engine works. Nine commands, each tested on real AI tools: check (every tool's rules, secrets included, each backed by a dated source), tokens (what your skills cost on every message, against each tool's budget), trigger (does the right skill fire, in Claude Code, Codex and Antigravity), package (one source into every tool's manifest, and Codex's "only when asked" file), repo (a new plugin repository that passes every tool's validator), eval and bench (with and without the skill, graded, with a review page), install (each tool's own validator, then the plugin loaded and run), overlap (proof nothing was copied), and release with changelog.
• Plugins, not only skills. Three more skills in the specification: /skill plugin for a whole plugin, /skill repo for the repository around it, /skill release for an update that reaches every user. 23 skills in all.
• A FAQ that answers what goes wrong, with what we measured: why skills stop triggering, how many fit, fields that silently do nothing, running only when asked, every tool, and safety and cost.

Added

• skill-learn: a correction or lesson applied to every skill it touches, sorted by how sure it is, rule-or-memory asked for each, a diff per file for your yes, versions bumped, and the affected evals run again.
• skill-compete: a skill you point at taken apart, ours built from scratch, both run on the same evals, its licence read, and a check that none of its text came across.
• skill-create from this chat, a screen recording or a pasted brief, citing where each rule came from and tested with and without before you see it. skill-check gains portability: one tool's names or memory paths, and macOS-only commands with no Linux fallback. skill-update watches every skill you competed with.
• How Skill Builder proves itself: the same tasks against a plain AI and other skill-creation tools, measured. 20 skills, 7 agents.

Changed

• The title, the README's first line, the GitHub description and every manifest now say the same thing in the same words: a plugin and skills for Claude Code, Codex and AI agents.

Added

• Antigravity: install steps for the Antigravity CLI and app, and Antigravity named wherever the README lists the AI tools it works in.
• The README, which is the specification for the first release: every skill and agent, how each one starts, the evidence rules, and install steps for Claude Code, Codex, Gemini CLI, Cursor and the skills CLI.
• Every AI tool's manifest, built from product.json by scripts/manifests/build.mjs and checked in CI, with no MCP server, since there is nothing to connect.
• The house files: licence, security policy, contributing notes, AGENTS.md and the GitHub topics.