Research, build, learn

As models evolve, we have to evolve with them.

Each model has a different purpose. I’m not going to ask GPT-5.x to generate graphics for me. It can, but I’ll probably go to Nano Banana or something like it.

I’ve been trying to share more examples of how I use AI day to day, mainly because I wish someone had shown me this a year ago. And maybe someone out there has a better idea.

Research

I’m in claude.ai, using a mix of Sonnet 4.x, Opus 4.x and Sonnet 5 on low effort. This is where I have no idea what I want to do yet, or I’m just riffing with AI. Out of laziness, my output style setup here is pretty bare. If I’m feeling crazy, or I really need help, I’ll jump into Fable.

Even so, I try not to start a session unless I have a pretty solid idea of where I want to go.

Building

I’m in the Claude Code CLI inside WSL2. I’m trying to make the move to cloud-only with Cowork, but I can’t let go of Claude Code inside VS Code. It’s configured to Sonnet 5 on high effort, with Advisor mode on (Opus 5, medium), and the output style set to concise.

I’m not chatting with it, or at least I’m trying not to. I’ve got a series of workflows built as plugins: engineering, architecture, and others focused on operating profiles that output assets. For example, one plugin’s entire job is to research an outcome and output a JSON file structured as an example for other prompts.

If I have to ask it to fix something more than two or three times, my workflow is missing something.

An average run is a loop of about 30 minutes before I have to review something.

I’m still not the best with worktrees and multi-session hand-offs. I’m still wrapping my head around the new session-to-session handlers.

Metrics, telemetry and value

This is where it gets interesting. I’m now using Opus 5 on extra-high effort to analyse the outputs of my building phase, again through plugins. Its focus is to read all the assets generated and give feedback on what could have gone better.

That feedback becomes a context folder of things that went wrong or could have gone better, which I inject into a lessons-learned folder for my next agent runs.

It’s not foolproof yet, and I’m sure there’s a better way to do it. Believe it or not, I’m doing this one inside Cowork. And it’s making my agent runs faster and, well, cheaper.

So, each phase

1. Research

   Who: claude.aiSonnet 4.xOpus 4.xSonnet 5Fable

   I’m chatting back and forth while the building agents run.

2. Building

   Who: Claude CodeSonnet 5Opus 5 advisorPlugins

   If I’m chatting here, it’s to go and fix my workflows.

3. Learn

   Who: CoworkOpus 5Lessons learned

   Working out how to make the building more efficient, and get much more value out of it.

Final thought

I’ll go deeper on my research phase in another post. I think it’s important to realise that’s where the hard work actually is: you’re going from ambiguity to reality.