Template repository · use with Claude, Codex, or any AI assistant

The Harness Engineering Stack

A pattern language for running the full software development lifecycle with AI assistants — as a team of one.

PLANspec + plan doc
─▶
BUILDTDD, own worktree
─▶
VERIFYone-command gate
─▶
SHIPCI/CD + release
─▶
LEARNhuman QA + memory

One thread runs through it all: spec → plan → commit → gate → release → QA charter — walkable in both directions, for every shipped feature.

The premise

The model in the loop will change. The harness is what compounds.

You are one person. Your “engineering team” is you plus one or more AI assistants — and the thing that makes that work is not the model. It is the harness: the repo conventions, commands, gates, and documents that let any assistant pick up work, verify its own output, and ship, without you re-explaining the project every session.

A solo developer’s constraint is no longer typing speed; it is verified attention — how much shipped change you can actually stand behind. Each pattern here converts something you would have to remember, repeat, or manually check into something the repository enforces — and add-on layers extend the same loop to corporate trackers, security, privacy, and on-device debugging.

The patterns

16 patterns · 7 core layers
II

Plan

What are we building next, exactly?

  1. 05Plan as Reviewable Artifact
III

Build & Verify

How does the assistant know it’s done?

  1. 06One-Command Interface
  2. 07Test-First Loop
  3. 08Verification Gate
IV

Integrate

How does work land safely — even with many agents at once?

  1. 09Small PRs on Trunk
  2. 10One Worktree per Agent
  3. 11CI Mirrors Local
V

Release

How does it reach users?

  1. 12Release from the Ledger
VI

Human QA

What can only a human judge?

  1. 13Human QA Charter
VII

Memory & Orientation

What persists between sessions, where does everything live, and who is working on what right now?

  1. 14Memory Index and Leaves
  2. 15Session Ledger
  3. 16Annotated Sitemap

The patterns (ADD-ON)

10 patterns · 6 add-on layers
VIII

Corporate

add-on

How does work flow in from the org’s tracker, and how much autonomy does each ticket get?

  1. 17Ticket Bridge
  2. 18Automation Tier Policy
  3. 19Evidence Pack
IX

Security & Privacy

add-on

How does security knowledge become permanent, and what must never leave your control?

  1. 20Security Learnings Become Tests
  2. 21The Confidentiality Line
X

Device Repro

add-on

What happens when a bug only exists on a physical device?

  1. 22Paired Device Repro
XI

Code Strengthening

add-on

How do unfinished features ship safely, and how do old clients keep working?

  1. 23Environment-Matrix Feature Flags
  2. 24Versioned Backend Contracts
XII

Image Analysis

add-on

How does an assistant see details instead of the gist?

  1. 25Slice-and-Compose Image Analysis
XIII

Onboarding

add-on

How do juniors become seniors when AI does junior-shaped work?

  1. 26Assisted Apprenticeship

Start in three moves

Take the template

Click Use this template on GitHub. You get the module skeleton — backend/, web/, mobile-app/, webservices/ — and the full pattern language in the README.

Point any assistant at it

Open the repo in Claude Code, Codex, Cursor — whatever you use. Paste the kickstart prompt from the README.

It scaffolds the harness

Constitution, docs tree, justfile verbs, CI gate, memory index, session ledger, sitemap — then it stops, and feature work begins with a spec you approve.