Build the Harness Out (Weeks 2 to 6)

Part 7 of the field-guide companion to Builder-Leader. Part 6 was the first Saturday: install, bootstrap, first rewrite, first skill, first hook. This is the next four weeks, where the setup goes from "I tried it" to "I work through it." Four moves, each spread across a week, each ending in real work.

Week 2: notice what you do twice

You don't plan a skill library. You discover it. Planned skills mostly fail because they're built for imagined work. Noticed skills mostly succeed because the need already existed.

Open a file called twice.md. For five working days, every time you catch yourself typing a similar prompt a second or third time (a status-note framing, a summarize request, a rewrite pattern), paste it in with a one-line note. On Friday, look for a cluster of three or more. Turn the cluster into a skill, generalize the prompt, add one line on when to use it, and run it on real work Monday.

A skill file with your name on it that did something useful on a Monday task: that's the week.side quest

Week 3: memory without overengineering

The trap here is building a memory system more elaborate than the work justifies. Avoid it with the minimum that actually compounds.

Split your CLAUDE.md in two. Keep the stable stuff (who you are, what you work on, tone) in CLAUDE.md. Put what's true this week (active workstreams, open questions, people in play) in CLAUDE-current.md. Point the system at both. At week's end, update the current file with whatever shifted, and leave CLAUDE.md alone unless something genuinely stable changed.

Two files, each shorter than you think they should be. The best operators describe their month-ten setup as looking a lot like their month-three setup: small, deliberate, quiet. Quiet is the compliment. It means the rig stopped being a project and became a tool.side quest

Week 4: spawn one agent

Skills you run synchronously. Agents handle a specialized sub-task you can hand off. The line is easy to cross too early, so the trigger is "recurring sub-task I actually do often," not "could I automate this?"

Pick the sub-task (pre-read summaries, follow-up coordination, first-pass review of someone else's doc). Define a subagent in .claude/agents/<name>.md with one job, one input format, one output format. Use it on a real instance this week, then again two days later. If the second use feels worse than doing it by hand, the agent is wrong: re-scope or delete. If it's faster and the output is usable, it cleared the bar.

Week 6: ship something

The first five weeks have to end in a shipped artifact. Not because shipping is the point, but because every operator who stopped short of shipping in this window retreated to pre-harness habits and stalled.

Do one real piece of work through the harness, end to end, for an actual audience: a team, a peer, a boss, a customer. At least as good as what you'd have produced without it. Note the time, and note one specific thing the harness did that you wouldn't have done yourself or would have done worse.

A piece of work delivered to someone expecting it, built through the harness, with one sentence on what the harness contributed. That's the threshold that turns "I'm trying Claude Code" into "I work through Claude Code now."

What you are now

A skill, a two-file memory, an agent, and one shipped thing. That's the definition: you're a Group-3 operator, the quiet subset that doesn't just use these systems but builds the structure around them. You got there in about six weeks, the way everyone who gets there does, by doing the reps rather than reading about them.

The full Chapter 7 field guide is in the public repo: field-guide/07-building-your-personal-harness.md. The book is at builder-leader.com.

Next: what changes when a second person joins you, and the two-person phase nobody should skip.