Chapter 6 · Capstone: Build Your Harness · Lesson 6.1
Your Starter Harness
The win: the minimum viable harness you can stand up today - one piece borrowed from each earlier chapter.
- 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
You already have all the parts
Every earlier chapter handed you one piece of a working harness - the scaffolding you wrap around a coding agent. This lesson does the assembly. The goal is not the biggest harness you can imagine; it is the smallest one worth having, standing up on a real project today.
The rule that governs the whole build is Lesson 1.1's: the ratchet - a mechanism that only ever tightens. You start with almost nothing and add a piece only after a real failure earns it. So do not build it all up front. Stand up the five slots below mostly empty, then let real runs fill them.
The five-piece starter kit
Each piece is deliberately tiny. Each one traces back to a lesson you have already done:
- A rules file with 3-5 earned lines, or empty to start - the rules file the agent reads every turn (Lesson 1.2). Empty is fine; unearned lines just spend attention.
- One hook - a small script that runs your tests automatically after each edit, so a break is caught 100% of the time, not 70-90% (hooks, Lesson 1.3).
- A clean git branch and a safe place to run - a fresh branch plus a sandbox, an isolated spot to run code without risking your real machine (Lesson 3.3).
- A tiny eval set of 3 golden tasks - a fixed eval set of real tasks with known-good outcomes, so you can tell if a harness change actually helped (Lesson 4.2).
- A review habit - one fresh-session second pass over the work, a cross-agent review that catches what the first pass missed (Lesson 1.6).
Why start this small
The temptation is to front-load: fifty rules, ten hooks, a rich config, all before the agent has done a single task. That harness rots. Half those rules were never earned, and unearned scaffolding just competes for the model's attention and goes stale on the next model release.
An empty-but-real harness that you ratchet beats a big speculative one every time. Osmani's lean-and-earned discipline is the same idea pointed at a fresh project: five slots, each nearly empty, each ready to tighten the moment a run fails.
Check yourself
The right way to start is -
Stand up the five slots mostly empty, then let real failures earn each addition. A big speculative harness rots; an empty-but-real one that ratchets keeps improving.
The one starter hook should -
A hook fires deterministically every time, so a broken test is caught 100% of the time rather than the ~70-90% you get from an instruction alone.
Your tiny eval set holds -
A small fixed set of real tasks with known-good outcomes lets you measure whether a harness change actually helped - your own benchmark, run after every change.
Do this now (15 min)
Pick one real project and stand up the five-piece kit:
- Create the rules file (
CLAUDE.md/AGENTS.md) - even near-empty. - Add one hook that runs your tests after an edit.
- Make a clean git branch to work on.
- Write down 3 golden tasks with clear done-when outcomes.
- Commit to doing one fresh-session review pass.
That is it. You now have a real harness - small, honest, and ready to ratchet as the project teaches you what to add.
Go deeper
Primary source (read this): Addy Osmani - Agent Harness Engineering. The source of the lean, earned discipline this kit is built on.
Wisdom (test it on people): the HumanLayer community - a good place to have your starter harness pressure-tested by people running the same loop.