addyosmani/agent-skills

The problem

AI coding agents are fast at producing code and weak at engineering process. Ask for a feature and you get a plausible diff, often without a spec, without tests that were written before the implementation, and without a review pass. Each session you re-explain the same expectations, and output quality drifts between agents and between team members. This pack addresses that by shipping the workflows themselves as installable skills, so the process travels with the agent instead of living in someone's head.

What's different

The usual approach is a long AGENTS.md or rules file, or a custom prompt per task. Those are prose instructions the model may or may not follow; they also tend to be one flat document. Here the process is split into 25 discrete, named skills, each a structured workflow with steps, verification gates and anti-rationalization tables, plus 9 slash commands mapped onto lifecycle stages from /spec to /ship. Skills also auto-activate from context — designing an interface pulls in the API design skill. The other distinguishing piece is /build auto, which lets the agent plan and implement a whole approved plan autonomously while still committing task by task and stopping on failures. That is a narrower, more opinionated bet than a general agent framework: this is a methodology, not a runtime.

Who it's for

Developers already using a terminal or IDE agent (Claude Code, Cursor, Codex, Gemini CLI, Copilot, and dozens more via the skills CLI) who want enforced discipline rather than raw generation speed. It suits teams that care about TDD and review gates and want those encoded once. It is less useful if you only use chat-style AI assistance, if you work in a language or stack where the lifecycle stages don't map cleanly, or if you want your agent to invent its own process — the pack is deliberately prescriptive.

Cost of adoption

Runtime is just your existing agent; the content is plain Markdown. Installation varies sharply by host: some are one command, others require syncing folders or pasting content into rules files, and slash commands are Claude Code–specific — in Copilot you invoke skills by name instead. A known portability gap: installing a single skill via npx copies only skills/<name>/ and not the shared references/ checklists, so per-skill installs lose supplementary material (tracked in issue #361). MIT licensed, so no legal friction. Expect to spend real time reading the adoption guide before rolling it out to an established codebase.

Verdict

The project is around eight months old and sits at roughly 110k stars with about 11.7k gained in a single tracked month, which is a steep slope for a repository made of Markdown — momentum is clearly rising, but a large share of that interest may be curiosity rather than sustained use. External discussion is thin: the top Hacker News thread drew 2 points and no comments, so community validation is mostly star-based. Treat it as a well-curated, actively maintained methodology pack rather than a proven standard. Worth trying on one repository if agent output discipline is a real pain; judge it by whether your diffs start arriving with tests and fewer review comments, not by the star count.