Levelling up my Claude Code skills

Published on 3 July 2026

Until fairly recently my workflow was:

  1. write a plan with an agent
  2. agent implement plan and ask it to spin up sub agents where applicable
  3. I then code review and verify the change (manually test it)

This was OK but I found I often spent my time babysitting the agent and being very involved in the work still.

I found the concept of developers with multiple Claude Code tabs open hard to wrap my head around! So I recently challenged myself to push a bit harder and lean into it more.

I’m starting to feel (at least) like a “power user” with multiple Claude Code tabs going at once

What changed

A couple of things nudged me in this direction:

  • I started poking around the “How Boris Uses Claude Code” resource as a source of inspiration to see what those who know more than me are doing.
  • I started using Superpowers which immediately produced better plans, specs, implementation plans, etc.

However I have found it pretty token hungry and there’s a fine balance where the amount of planning and pre implementation ceremony is worth it. I still tend to use it but I can see myself in the future having my own system or framework to operate with.

The Superpowers plugin has helped me with a significant migration in the last week. It’s been running largely in the background with me occasionally checking up on it. It also wrote a manual testing plan for it to work through using my browser. This is the kind of work that would never have been prioritised pre-agentic-coding.

Starting to build my own agentic framework

As a result, I’ve started putting together a few things to enforce how I want my agents to operate:

1) Encode “taste” as skills

The Claude Ruby Quality plugin provides a couple of skills that detect and remedy common code smells and test smells, by encoding canonical Ruby Science code smells and their fixes, and xUnit test patterns.

2) Turn milestones into commands

Commands that I run at common milestones: for example an /open-pr command that chains Anthropic produced engineering skills. This means when I check a PR I can keep my analysis high level and be confident it’s been code reviewed, refactored and verified. I then sometimes add a few comments and get those addressed before releasing.

I even have commands that can trigger our CI pipeline so I can stay in Claude Code.

3) Write a CLAUDE.md that enforces my defaults

My CLAUDE.md imparts some stylistic taste I want agents to adopt. I have taste when it comes to Rails apps that I want to enforce, and I find if left unattended, you end up with a /services directory containing a collection of random objects.

It enforces a 37Signals-style approach — /models only and ViewComponent by default — and prefers server-driven UI changes where possible.

My “verify-review” command

Once an agent has implemented something, I have a verify-review command I run before opening a PR:

---
description: Run /code-review --fix, then verify the result in the browser with /verify
---

Do the following in order, completing each step before moving to the next:

1. Invoke the `code-review` skill with `--fix` to review the current diff and apply the findings to the working tree.

2. Once the fixes are applied, invoke the `verify` skill to confirm the change actually works by running the app and observing its behavior in the browser via the `claude-in-chrome` skill (load the `mcp__claude-in-chrome__*` tools through ToolSearch as that skill directs).

If step 1 applies no fixes (clean review), still run step 2 to verify the existing change. If verification surfaces a regression, report it clearly rather than silently re-editing.

I’ve found the code review particularly effective. Until recently I have been code reviewing and I now trust this skill enough that I only need to pay attention to the high level; I don’t need to go through every line critically.

The big thing for me is that the code-review skill spins up sub agents to review the diff from different angles — security, behaviour, and more. Having several independent reviewers each looking for something specific is really powerful.

It’s commonly known that agents write superfluous code and often disobey YAGNI liberally. /simplify guards specifically against this, trimming back anything that didn’t need to be written.

The verification using Claude in Chrome is helpful too because I can easily see what’s being tested, or let it run in the background.

Going all in on Claude Code (for now)

I’ve previously said I want to stay tool agnostic whilst things are changing so quickly; I bounce between Codex, Claude Code and Cursor. However recently I’ve given up my Codex and Cursor licences and gone all in on Claude Code.

This has unlocked a step change in productivity – and I think the concepts transfer neatly to other tools.

I’ve also heard a lot about “loops” being the new thing – that’s the next area I want to dig into.