Chapter 6 · Capstone: Build Your Harness · Lesson 6.3
The First-Week Ratchet Plan
The win: a concrete five-day plan that turns the harness discipline into a habit, one small rep at a time.
- Chapter 0 · Sprint Zero
- Chapter 1 · The ratchet & the practice loop
- Chapter 2 · Spec-driven development in depth
- Chapter 3 · Scaling & trusting the harness
- Chapter 4 · Measuring & evolving the harness
- Chapter 5 · Multi-agent & team harnesses
- 6.1 · Your starter harness
- 6.2 · One feature, end to end
- 6.3 · The first-week ratchet plan
- 6.4 · Adapt, don't copy
- 6.5 · Staying current & your capstone
Small reps beat a big-bang setup
You now have a starter harness and one feature run through it. The trap at this point is to block out a weekend, write a huge rules file, wire up every hook you can think of, and call the harness "done". That is the exact opposite of what worked all course. The ratchet only tightens after a real failure, and it compounds: a week of tiny, earned additions leaves you with a real harness you understand and trust, where a big-bang setup leaves you with a pile of guesses you'll never prune.
So the goal of week one is not a finished harness. It's a habit. Five days, one small move each, and by Friday the practice runs itself. Here is the plan.
- Day 1 - create and commit the rules file. Add nothing speculative; start near-empty (Lesson 1.2).
- Day 2 - add your first hook: run the tests after edits (Lesson 1.3).
- Day 3 - write 3 golden tasks, your eval set (Lesson 4.2).
- Day 4 - do one cross-agent review pass on real work (Lesson 1.6).
- Day 5 - review the week's failures, ratchet the top one into a rule or hook, and prune one unearned line (fights drift, Lesson 5.5).
Each day is one 15-minute move, and each one stands alone. Miss a day and you pick it back up the next; nothing later depends on doing everything perfectly. The point is to touch the harness five times, not to finish it.
After week one: keep only Day 5
Days 1 through 4 are one-time setup. Day 5 is the practice. From week two on, repeat Day 5 every week: review the worst failure, ratchet it into a rule or hook, and prune one stale line. That single weekly habit - tighten the worst failure, drop one unearned rule - is the entire discipline, sustained. It's the compound loop pointed at your own harness: the model doesn't get smarter, the harness does.
Keep it honest
You will be tempted, probably on Day 1, to front-load a huge rules file "just in case". Don't. Every line still has to be earned by a real failure, exactly as it has all course - an unearned line just competes for the model's attention and dilutes the rules that matter. A near-empty rules file you'll grow honestly beats a bloated one you'll never trust or prune.
Check yourself
A real harness is best built -
A harness is grown in reps, not built in a day. Tiny earned additions compound; a weekend big-bang leaves guesses you never prune. The five-day plan is one small move per day.
After week one, the sustained habit is -
Days 1-4 are one-time setup; Day 5 is the practice. Weekly: ratchet the worst failure into a rule or hook, and prune one stale line. That single habit is the whole discipline.
On Day 1 the rules file should -
Front-loading is the trap. Every line must be earned by a real failure; unearned lines dilute the ones that matter. Commit a near-empty file and grow it honestly over the week.
Do this now (2 min)
Put the five days on your calendar - one 15-minute slot each. Then do Day 1 right now: create the rules file for your project and commit it, even if it's nearly empty. That commit is the start of the ratchet; everything else this week builds on it.
Go deeper
Primary source (read this): Addy Osmani - Agent Harness Engineering. The case for a harness earned failure by failure, not built up front.
Wisdom (test it on people): the HumanLayer community - a good place to keep your weekly Day 5 pruning honest with people doing the same practice.