Harness Creator — make your coding agent reliable across sessions

The single biggest upgrade to working with a coding agent is not a better prompt — it's a harness: a small set of files in your repo that tell the agent how to start, what "done" means, and how to pick up where the last session left off. Without one, every session starts from amnesia and "done" means "the agent got tired." With one, work survives sessions, scope stops drifting, and nothing is called finished until a verification command actually passes.

Harness Creator is a skill that builds and audits these harnesses for you. It's a third-party skill by walkinglabs — install it from the source (one command below); this page teaches you what it does and how to get the most from it.

Install (one command)

npx skills add walkinglabs/learn-harness-engineering --skill harness-creator

That's the skills.sh installer — it drops the skill into your .claude/skills/ so Claude Code loads it when harness work comes up. (New to skills? Read /skills-guide first — five minutes.)

Check it worked: ask Claude Code to "audit this repo's harness" — it should invoke the skill and score five subsystems rather than improvising.

What a harness actually is — five subsystems

Subsystem	The file	What it prevents
Instructions	`AGENTS.md`	agent starts differently every session
State	`feature_list.json`, `progress.md`	"what was I doing?" amnesia
Verification	`init.sh` (the gate)	"done" without passing tests
Scope	feature dependencies + done-criteria	half-finished everything
Lifecycle	`session-handoff.md`	next session can't resume

The skill scaffolds these, validates existing ones with a 100-point score, and tells you which subsystem is your bottleneck.

See a real one

This hub hosts a working example the skill built and scored 100/100: /build-your-own-x-harness — the famous tutorial index wrapped in a full harness (read its AGENTS.md, feature_list.json, and init.sh to see every subsystem in place). Clone it, then run the skill against your own project and compare.

How to use it well

Point it at a real project, not an empty folder — it inspects what exists (manifests, test commands, docs) before scaffolding.
Replace the placeholder features immediately. A harness with fake features is theater; write the 5-8 real features with real verify commands.
The verify gate is the product. If init.sh doesn't actually fail when the code is broken, the whole harness is decoration.
Audit before you extend — on an existing project, run the validator first and fix the lowest-scoring subsystem before adding anything fancy.

Why this matters for everything else on this hub

Every serious project here follows harness discipline — verification gates, state files, evidence before "done." The agents ladder (/ai-agents-setup) gets you agents; this skill is how you make them accountable.