I catch myself mid-keystroke. Claude Code already has the task. I'm about to tell it how I'd solve it — steps, caveats, the path I'd take if I were typing.
I stop. Leave the ask. Let it run.
That's the rep I'm practising. Not another harness. Not overnight review. Getting out of the way so the model can do the work I'd normally do myself.
Treat it like a brilliant developer on day one. Sharp. Fast. Almost no memory of how you work, because the short-term memory resets every session. That isn't a failure mode. It's why the first prompt carries more weight than the tenth nudge mid-flight.
People say agents aren't consistent. Fair enough. Human consistency comes from reps, intuition, the aha after you've seen the same shape fifty times. The model doesn't inherit those across cold starts. You do. Your job is to put enough of that into the ask that it can act like it was already in the room — then stop dissecting while it works.
I started small so I could feel the ask. One-liners first: rename a helper, find why a test flakes, explain a path I'd usually dig myself. Same instinct as coaching someone new — one clear task, see what comes back, tighten the next brief. Break the big thing until each chunk has that clarity. One task isn't a productivity aesthetic. It's how the short-memory teammate stays oriented.
When the output's soft, I don't always stay in the thread and poke. Often I revert, edit the initial prompt, and run again. The prompt does more than the mid-course correction. Give it freedom to work — not a leash that yanks it onto your preferred path, and not so little that it wanders into a tangent you never meant.
Sometimes getting out of the way means asking for the work before the typing: research this, find the bug, explain the system, recommend an approach I'd actually follow. Let it show its homework. That's still me steering — just earlier, where a better ask compounds.
The craft lives in a file. I keep a PROMPT.md next to the CLI (command-line interface) session: what we're doing, what good looks like, the constraints that matter. When the context gets mushy I reset. Fresh session. "Work on PROMPT.md." The thread starts clean. The intent survives in the file. You're not abandoning the work. You're refusing to hand-hold through a conversation that already forgot the sharp edges.
Anything I'd say again goes into AGENTS.md — naming, what not to touch, the review bar. CLAUDE.md points at it so the harness finds the house rules without me pasting them every cold start. Nested directories get their own notes where the local shape differs. Reps become documents. The aha from last Tuesday doesn't have to be rediscovered every morning.
Context fills and the model gets dumber — not suddenly stupid, softer around the edges you cared about. Don't cram. Hand off. Spawn a subagent. Write a short handover: what we know, what's open, where it can go wrong. Same bet as refusing to sit on the suitcase in the agent factory — working memory stays small on purpose. The next agent reads what it needs.
My notes for this piece cut off mid-plan. I had Research and Execute, and a blank where Plan should sit.
That blank is the point.
Research is the homework. Execute is change the code. Plan is the middle I kept skipping — a short pass that names the shape before anyone touches files. Not a tutorial checklist. Just: after you know enough, before the patch, put the path in the prompt (or in PROMPT.md) so the short-memory teammate isn't inventing the route while walking it.
I used to fill that gap myself, one correction at a time. Getting out of the way means putting the plan upstream so the work can run without me.
Task. See the output. Help it get better.
Don't hand it your solution. Leave the ask. Then get out of the way. That's the skill.