Small Steps

Drive AI agents through a million lines of code you did not write. Recon, surgical navigation, codemods, characterization tests, and a toolkit of four reusable skills you keep.
We have the instructions, the skills, and the map. We know what needs to happen. So let us do it.
I am going to remove Copilot two different ways, on two branches, and Git is going to save me if either one goes pear-shaped. The first way is small steps: no diff bigger than 1,000 lines, so I can review every piece of the removal as it happens. The second way, in a later episode, is the big lift. Kick it off, go to bed, wake up to a 25,000 line change and ask myself whether that is what I wanted.
Small steps first, and that needs a tight plan.
I switch back to Opus 5 for this. Sonnet could probably do it, our instructions are bolted down pretty hard, but I want the lateral thinking Opus brings when something goes sideways mid-removal. Then the prompt: help me remove Copilot from this application entirely. I want to run VS Code and not see a single trace of it. You have a map and you know where the Copilot code lives. Do it in small steps so I can review at every point. Create a branch, and every task, every step, has to be a 1,000 line change or less. Lay the plan out in docs/copilot-removal.md and tell me what you need.
A note from the edit bay: I never said the word PR. I said I wanted to review 1,000 lines or less, so what Claude built was commits of 1,000 lines or less. That worked fine, since I could diff each commit and walk through it, and you will see that in the next episode. But if you were wondering why there are no pull requests, that is why.
The plan took four or five minutes. 31 steps, eight phases, three checkpoints, plus a few questions for me. The one I should have seen coming: there are two kinds of removal here. Deleting a whole file, which I can review by manifest, and editing a file, which I have to read by hand. Fine.
Then it asked about telemetry. VS Code uses a lot of it and I do not need any of it, so out it goes too. I tell it the plan looks good, add telemetry to the removal, and make the plan a task list with checkboxes so I can watch progress tick by. And then I said "let's go."
I am not being naive about what happens next. The people who wrote VS Code are high-powered engineers who know how to build malleable software, so in theory this comes out clean. In practice I expect everything to break the first time I try to run it, and at that point I refactor and fix until it is stable. Same goes for the big lift later, assuming I do not run out of tokens.
Which brings me to the part I love. I had to stop it. It had started executing the work. When I asked why, the answer was, more or less, "you said let's go." And I did. Opus was excited. Three commits were already on the branch.
I need to be more careful about what I say to these things. I am also, plainly, happy that work is happening. The removal doc updated itself to "Copilot + Telemetry Removal," the rules we agreed on are in it, and the checkboxes are ticking off through phase one.
The question for the next episode: will it actually stop at the review checkpoint, or plow straight through?
Drive AI agents through a million lines of code you did not write. Recon, surgical navigation, codemods, characterization tests, and a toolkit of four reusable skills you keep.