Harness Engineering Starter Kit
Copy-ready templates for repository instructions, persistent progress, bounded features, evidence-based completion, and browser verification.
Updated 18 Aug 2026· no account required
Start with three files, not nine.
Copy `AGENTS.md`, `PROGRESS.md`, and `feature-list.json` first. Replace the placeholders, define one bounded task, and name the command or browser procedure that proves it works. Add the audit and specialized checklists when a repeated failure shows where the harness is weak.
project/
|-- AGENTS.md
|-- PROGRESS.md
|-- feature-list.json
|-- src/
`-- tests/What is inside the starter kit?
AGENTS.md
Project map, standard commands, working procedure, boundaries, and definition of done.
PROGRESS.md
Verified work, incomplete work, decisions, evidence, and the next bounded action.
feature-list.json
Structured scope, non-goals, acceptance criteria, verification, and evidence.
harness-audit.md
A five-subsystem scorecard for instructions, tools, environment, state, and feedback.
definition-of-done.md
Completion checks covering outcomes, code quality, runtime behavior, safety, and handoff.
browser-verification.md
A repeatable browser procedure with console, network, keyboard, and payment-flow checks.
repository-structures.md
Minimal, growing, and multi-package layouts with guidance on when each is warranted.
payment-flow-task.md
A worked task specification that tests Pay now through success, decline, and duplicate clicks.
README.md
The adoption order and a concise explanation of every file in the kit.
Use the kit as a feedback loop.
- 01
Map the repository
Fill in the purpose, exact toolchain, important directories, standard commands, and hard boundaries in AGENTS.md.
- 02
Bound one task
Write the outcome, allowed scope, non-goals, acceptance criteria, and required evidence in feature-list.json.
- 03
Execute and observe
Let the agent work through the real project tools. Use tests, browser behavior, network requests, logs, and the diff as feedback.
- 04
Verify before completion
Run the selected definition-of-done checks and record evidence. A created file or rendered page is not enough.
- 05
Leave a clean handoff
Update PROGRESS.md with verified work, incomplete work, decisions, failures, and one next bounded action.
Measure whether the basic loop removes a real failure before adding reviewers, schedulers, or multi-agent coordination.