Six Months In: Building a Thinking Partner · Post 1 of 4

I Wanted a Thinking Partner, Not a Tool

Pete Griffin, Digital VisionOctober 2026

My business had gotten stale. When I came back to building this year, AI was the thing that made me want to dream about it again.

The ambition was specific. I didn't want a faster way to type code. I wanted to make Claude the perfect thinking partner: something to plan with, argue with, and build alongside.

So I went all in.

The sprint

In spring I ran a build sprint and tested nearly everything Claude Code can do. Plugins, skills, specialist agents, swarms, workflows. The install records show how it went.

It was hard work and I enjoyed it. You can't know in advance which parts of a new way of working will matter. The only way to find out is to try them properly.

The count

Then the sprint ended. The version ones were released, the apps went into review, and the work got quieter.

In October I asked Claude a question. Three months on, what does day-to-day work still use? It read the sessions saved on my laptop and counted.

I'll be exact about what that covers. One laptop, 22 July to 10 October, 33 working sessions, 640 messages from me. The sprint itself isn't in it, and neither is my main computer, where most of the testing happened. It's a picture of what stays once the experimenting stops.

Two skills did most of the skill work. One questions my plan before I build. One polishes interface design. 21 of 32 agent runs were reviewers, checking work that was already done.

What that told me

It didn't tell me the sprint was wasted. The sprint is how I found the eight.

What it told me is what a thinking partner is actually made of. Something that questions the plan. Something that checks the work. A memory of how I like things done. The rest was scaffolding I needed in order to find those.

Putting the scaffolding away

Scaffolding has a cost if you leave it up. Before I type a word, every session loads a description of everything installed. In July that starting load was about 60,000 tokens. By October it was about 100,000.

So I switched off 41 of the 49 plugins. It was one edit, and any of them can come back the day a project needs it.

The rule: dream big and test widely. Then keep what the ordinary weeks use, and put the rest away until you need it.

The numbers in this series all come from one audit of my own Claude sessions. Read the full report →

Next in the series · Post 2 of 4
Ambition vs Reality: Five Things I Pictured, and What They Became