I run a one-person web agency and app studio. My business had gotten stale, and AI made me want to dream about it again. The ambition was to make Claude the perfect thinking partner. In spring I ran a build sprint and tested nearly everything Claude Code can do, mostly on my main computer. In October, three months after the sprint ended, I asked Claude to read the sessions saved on my laptop and tell me what day-to-day work still uses. This is what came back.
During the sprint I installed and tested a lot, and that is how I found what matters. Once it ended, a handful of habits did nearly all the day-to-day work: asking Claude to question my plan before building, having a second agent review the work, and writing my rules down so they carry into the next session. I switched most of the rest off this week. Any of it can be switched back on when a project needs it.
Dates come from plugin install records and the version history of my config.
Counts from 33 working sessions on one laptop, 22 July to 10 October, 640 messages from me. The sprint, and sessions on my main computer, are not included.
Skill runs. Bars share one scale (longest bar = 9 runs).
Agent runs. Reviewer agents are code review, security review, verification and critique (longest bar = 21 runs).
Carrying the whole testing kit is not free. Every session loads a description of everything installed before I type a word. That starting load was about 60,000 tokens in July and about 100,000 by October. Each new session now starts with 41 fewer plugins to describe; I will measure the saving over the coming weeks.
What I pictured for each thing, and what it became in the three months after the sprint. Reasons are taken from what I wrote in my own instructions at the time.
| Ambition | What it became |
|---|---|
Install everything, let Claude pick the right toolI wanted skills to remove thinking, with Claude choosing for me.Peter installs skills so he doesn't have to think about which one to use. |
Not in day-to-day use After the sprint, eight skills covered the work on my laptop. The rest stayed in the list and made every session heavier. A long menu did not make Claude reach for more of it. |
| Swarms of agents working on all my apps at onceI pictured one instruction fanning out across four apps in parallel. | Tested, then shelved I tested swarms during the sprint on my main computer. In the three months after, on the laptop, they had zero runs and the service was not starting there. Day-to-day work had not asked for it. |
Skills would run themselves once I trusted themMaybe after a few days or weeks I'll understand the powers I now have better, and you can just run them without consent. |
Partly One design skill runs automatically. Everything else still asks first, five months later. The reminder to revisit it "in late May" was never acted on. Asking first turned out to be what I want; I had changed my mind and not told the notes. |
| A team of specialists, each on the right-sized modelSmall model for lookups, large model for architecture, a dedicated agent for writing code. | Partly After the sprint, the reviewers are what stayed. On the laptop, 27 of 32 agent runs used whatever model I was already on. Claude mostly does the work itself and calls a second opinion at the end. |
| One command to release an appI wrote a release skill to replace five manual checks with one step. | Fixed for both machines I built it on my main computer. On my laptop it pointed at a folder that only exists on that other machine, and four of my own skills were dead links there. The audit found it and it now works from both. In the meantime the checks happened by hand, through the reviewer agents. |
| Claude remembers, so I never say it twiceA memory folder of notes: preferences, corrections, project facts. | Works, needs weeding 69 notes, loaded every session, still being added to. This is the only way Claude "learns" about me. But one note specified a colour my brand rules ban, one was five months stale, and the instructions pointed at a memory folder that did not exist on this machine. |
| A wiki that captures what each session learnedAutomatic, so knowledge builds up without effort. | Gone quiet Nothing has been added since 19 August. Anything described as automatic needs checking that it is still running. |
| Question the plan before buildingA skill that interviews me one question at a time before any real build. | Most used skill Nine runs. It is the cheapest thing in the setup and the one I would keep if I could keep only one. |
| Never accept "it works" without proofWritten rules: test it, show the evidence, or say it is unverified. | Stuck Two thirds of agent runs on the laptop were reviews or verification. Claude also opened real pages in a browser 486 times to check its own work. |
| Same setup on two computersOne synced folder for settings, skills and memory. | Mostly Settings and memory travel well. But 52 of 135 skill links on the laptop were broken, because they were written as addresses on the other machine. |
Not by itself. The model does not change between sessions. What carries over is what gets written into files: an instructions file and a folder of short notes. When I correct Claude and it saves a note, the next session starts with that note. That is real, and it adds up.
It also decays. Notes go stale, contradict each other, and point at things that have moved. Nobody tidies them unless you ask. Treat it like a shared handbook that needs an editor.
How this was measured. Claude read the session transcripts saved on my laptop and counted tool, skill and agent calls. Only sessions from 22 July onward were still on disk, so usage figures cover that window on that one machine. The spring build sprint and all sessions on my main computer, where most of the agent and workflow testing happened, are not in these numbers. Earlier dates come from plugin install records and the version history of my config folder.
Not measured: whether I correct Claude less often than I used to. A keyword count was too crude to trust, so it is left out. Client names and project details are removed.