What I learned building my own Claude Code plugin

For the last 3 months my main constraint with Claude has been me. Writing down requirements, testing output, approving commands. All these take time and babysitting that I couldn't devote to a side project. 2 weeks ago I redid my workflow and for the first time tokens became the bottleneck instead of me. Here's what I learned in the process:

My biggest takeaway is that if you've been procrastinating personalizing your workflow, stop what you're doing right now and invest a tiny bit of time to make it your own. I've been using GSD for most of the last 3 months and it's great. But even 2 hours of personalizing was a step change in how much I was able to get done.

The core insight for why workflow matters - it's all context management. Working with Claude is all about managing the Memento meme. It's a highly capable savant with an extremely short term memory and an ability to completely change its personality to match the needs of the task at hand. The best way to maximize performance is to keep context clean, clear, and focused on the specific job it's doing. Even a tiny bit of context pollution can drastically degrade performance.

The biggest unlock for me was taking tasks through a 5 stage workflow. Idea -> Backlog Ready Task -> Execution -> Verification -> Code Review. Each one requires a different kind of headspace and focus which is why using specialized agents is such a lift. If you're used to alternating between plan mode and execution you're already doing 2 and 3. Adding those three extra steps makes the whole process a lot more deterministic.

Idea extraction and task refinement are about taking ideas from your brain and turning them into Claude-ready artifacts. Claude is better at talking to Claude than you are. Put it into "understand what I want" mode first, then "plan how to build it" mode. What you get out are execution-ready artifacts it can actually run with. Plan mode is already 90% there but I found it valuable to split into two stages.

Executing those plans effectively was where the workflow really started paying off. Nothing particularly novel about it, but taking a task through separate execution, verification, and code review stages significantly improves output vs. having Claude tackle all three at once. It also catches the occasional sloppiness that can creep in when Claude has tunnel vision trying to get a task done.

The next thing that made a real difference was giving Claude a place to park things. You know the feeling when you give Claude a task, go get coffee expecting it done, and come back to a "would you like to proceed" permission request. Instead of blocking work, Claude moves on to the next task and adds an item to the human review queue for me. What's more, if it spots an issue unrelated to its task, instead of just ignoring it, it logs it as an observation that I can triage later.

Another big unblock was visual verification. This is something [@boris_cherny] talks a lot about, but giving Claude a verifiable outcome makes the result a lot more deterministic. I built specialized agents that use Playwright (web) and Maestro (mobile) to actually check the experience instead of assuming the code is right. As a result, when I do QA I'm not spending my time catching bugs, but focusing on how I want to improve the experience itself.

Most of the other steps I expected to help, but the one that surprised me most was a compounding step. This is one I learned from the team at Every and it's an extra step after an execution sprint has been completed to synthesize learnings and feed them back in as improvements. This added loop means the system is continuously improving.

Four buckets I get back from Claude every time:

• Immediate clean-up items
• Tasks we should add to the backlog
• Improvements to Claude.md and project documentation
• Improvements I should make to the workflow itself

I batch these so I don't look at them every sprint, but over time the benefits compound. That's why the Every team calls it compound engineering.

The net effect of all these changes is suddenly my progress took off. It took me 3 months for my first mobile app with Claude. It took me 1 week for my second. I switched from mostly having to sit at my computer managing Claude to watching it execute from my phone and occasionally nudging it along. Instead of single turn request and response, my Claude feed now looks like this:

If you want to do this yourself, get started with a simple prompt: "Build me a 5 stage execution workflow taking tasks through these five stages: Idea -> Backlog Ready Task -> Execution -> Verification -> Code Review. Each one should get its own specialized agent and slash command that I can reuse across projects. Ask me 3-5 clarifying questions before writing anything."

Or start with mine: github.com/kesteva/soloflow. You can install the plugin directly or just point Claude at it and take the pieces you want. It's MIT license for a reason and I did the exact same thing just telling Claude to look at GSD and Superpowers because those were the best models I saw when I was getting started. Don't wait until tomorrow, do it now. Thank me later.