Skip to content
ReviewRamble

Analysis Coding

Superpowers, Agent Skills, Pocock or ECC? Choosing a discipline for Claude Code

By Sheela ·

Coding agents guess. A better model guesses better, but it still guesses. The four most-starred skill packs for Claude Code all try to fix that the same way: by making the agent follow a process. Clarify, plan, test, review, then claim success.

They all ship that process as skills: folders of Markdown instructions that the agent loads when a task matches. So the real differences are about who drives the process, how much of the work it covers, and how much else gets installed.

The short answer

  • Hand off a big, vague task and come back to a reviewed result: Superpowers (4.7).
  • Take a feature from spec to production with the same steps every time: Addy Osmani’s Agent Skills (4.3).
  • Your agent keeps misunderstanding you and you don’t want a framework running your day: Matt Pocock’s skills (4.5).
  • You want hooks, memory, language-specific reviewers and a security scanner: a small profile of ECC (3.8).
  • You are fixing one well-understood bug: none of them. Process costs more than it saves on small, clear jobs.

Superpowers: the framework drives

Superpowers loads a small bootstrap at the start of every session that tells the agent to look for a matching skill before doing anything. Ask it to build something and it brainstorms with you, writes a plan of short tasks, and works through them with test-first development and a review after each task. It is the pack we would install first if you want an agent to plan and test before it codes.

The cost is pace. On small changes it is slow and uses a lot of tokens, and its hook-driven bootstrap can clash with other workflow plugins.

Agent Skills: the whole lifecycle

Addy Osmani’s pack has 25 skills and nine lifecycle commands that run from spec to ship, including browser testing and performance work that the other packs skip. By default there is a human checkpoint between phases.

It works best adopted whole. Installing single skills loses the shared checklists, and its /review command clashes with the one built into Claude Code.

Matt Pocock’s skills: you drive

These are small, single-purpose skills you call when you want them. The standout is /grill-me, which questions you one point at a time until you and the agent agree on what you are building. /grill-with-docs also keeps a glossary of your project’s terms so plans, code and conversation use the same names.

There is less automation, so the discipline is yours to supply, and each repo needs a one-time setup step.

ECC: the whole workshop

Everything Claude Code, now just ECC, is a large kit of agents, skills, hooks, rules, memory and a scanner for risky agent configuration. It is well maintained and has useful parts. Installed in full, though, it takes a lot of context and overlaps the other frameworks. Its own installer offers smaller profiles; use one. The project has been renamed and mirrored, so check you are installing the official package.

Don’t stack routers

Each of these has a bootstrap or meta-skill that decides which workflow to run. Install two and they compete: commands like /plan or /tdd get defined twice and the agent gets mixed instructions. Pick one primary pack. Borrowing a single skill from another, such as /grill-me alongside Superpowers, is fine.

Smaller options worth knowing

Every plugin that installs hooks runs commands on your machine. Read what you install, and prefer the official repository over mirrors.