Which agent-skills framework for which job

Addy Osmani published a clear comparison of agent-skills frameworks this week. It is worth reading on its own. If you are weighing these tools up, here is a short, practical map of what each one is built for. I have added Shipwright, the one I maintain, so the picture is complete.

One note before the list. These are not really rivals in the way a comparison table suggests. They optimise for different moments in the work. The useful question is not which one is best, but which one fits the job in front of you.

What each one is for

agent-skills (Addy Osmani). The full senior-engineering lifecycle expressed as skills, with one slash command per phase and a human checkpoint at each. It runs across many tools, including Claude Code, Cursor, Gemini and Copilot. A good fit when you want to walk a feature through every phase yourself and stay in the loop at each step.

Superpowers (obra). A complete methodology built around a disciplined build loop, with subagent isolation and git-worktree separation so parallel work does not collide. A good fit for long, autonomous, reasoning-heavy work where you want the agent to run for a while on its own.

Matt Pocock's skills. One expert's lean daily toolkit, sharp and low on ceremony. A good fit for a fast, battle-tested loop on TypeScript-style projects, where you want sensible defaults and not much overhead.

Shipwright. Keeps you in control of a codebase as it grows. Intent and specs are kept current and enforced on every change, so drift gets caught instead of quietly accumulating, and you can prove what you shipped at any point. It runs on Claude Code. A good fit for technical people who have moved past vibe coding and cannot afford to lose track of what their AI is building, yet do not have the time or the engineering background to run a full SDLC by hand.

They overlap, and that is fine

You do not have to marry one of these. You can cherry-pick individual skills across all of them and assemble something that fits how you work. The one thing worth avoiding, and this is Osmani's own caveat, is running two of them as simultaneous routers. They end up fighting over the same slash commands, and you get routing conflicts instead of help. Pick one to drive, and borrow skills from the rest.

A note on "runs on Claude Code"

Shipwright is in beta, and for now it runs on Claude Code only. It leans heavily on hooks, the part of the harness that can enforce a step rather than just suggest it, and it wires them across the whole lifecycle so every change gets checked against what you asked for. Several of these tools use hooks in some form, but Shipwright depends on a lot of them working together reliably, end to end. That is easiest to guarantee on one platform first. From there it gets ported to others that support hooks, such as Cursor or OpenAI Codex.

Where to start

If you want the source material, Osmani's comparison is the place to begin. If you want to see how the control-over-time idea plays out in practice, Shipwright is open source, and the code is on GitHub.

 

---

Sources:

• The Comparison of Agent Skills Frameworks - Addy Osmani - 2026 - https://github.com/addyosmani/agent-skills/blob/main/docs/comparison.md
• Hooks - Cursor Docs - 2026 - https://cursor.com/docs/hooks
• Cursor 1.7 Adds Hooks for Agent Lifecycle Control - InfoQ - October 2025 - https://www.infoq.com/news/2025/10/cursor-hooks/
• Hooks - Codex, OpenAI Developers - 2026 - https://developers.openai.com/codex/hooks