walkinglabs/learn-harness-engineering

What it claims

The course claims that agent reliability on real repositories comes from the environment around the model, not from the model or the prompt. Its headline evidence is an Anthropic experiment it cites: the same model (Opus 4.5) and the same prompt ("build a 2D retro game editor") gave different results depending on the harness.

• Without a harness: $9 spent over 20 minutes, producing something that didn't work.
• With a full planner + generator + evaluator harness: $200 spent over 6 hours, producing a playable game.

It also cites OpenAI reporting a "qualitative shift" from unreliable to reliable with Codex in a well-harnessed repo. Both results are second-hand. The course does not run its own measurements on this page.

How it works

The core prescription is a fixed per-session lifecycle driven by files in the repo. It is the part you would actually copy.

flowchart TD
  A[Read AGENTS.md / CLAUDE.md] --> B[Run init.sh]
  B --> C[Read progress log, feature_list.json, git log]
  C --> D[Pick ONE unfinished feature]
  D --> E[Implement]
  E --> F{Tests, lint, type-check pass?}
  F -- no --> E
  F -- yes --> G[Record evidence]
  G --> H[Update progress + feature list, note what's broken]
  H --> I[Commit only if safe to resume]

The design rules attached to this lifecycle:

• Instructions are split across files. The agent navigates them on demand ("progressive disclosure") rather than reading one giant file.
• Only a passing test suite counts as "done."
• The agent may not rewrite the feature list to hide unfinished work.

Later phases move beyond the single session. L13/P07 cover loops: goal loop, timer loop, maker-checker. L14/P08 cover graphs: fan-out/fan-in, rollback edges, human-in-the-loop.

How to use it

Minimal adoption. Drop four files from the Resource Library into your project root: AGENTS.md (or CLAUDE.md), init.sh, feature_list.json and claude-progress.md.

Despite its name, claude-progress.md is agent-agnostic. It works with Codex, OpenHands and others, but no agent updates it automatically. Your root instructions must tell the agent to read it at startup and update it before handoff.

Audit an existing repo. audit-harness.sh is pure shell and needs no Node.js. It exits 0 when all CRITICAL items pass, so it can serve as a CI gate:

curl -fsSL https://raw.githubusercontent.com/walkinglabs/learn-harness-engineering/main/tools/audit-harness.sh | bash -s -- /path/to/your/repo
# or after cloning
bash tools/audit-harness.sh /path/to/your/repo

The harness-creator skill ships a separate validate-harness.mjs.

Run the course locally. The docs site uses VitePress:

npm install
npm run docs:dev
npm run pdf:build   # PDFs to artifacts/pdfs/

Project code. Projects are an Electron + TypeScript + React app under projects/project-NN/, each with starter/ and solution/ directories. Each project's solution becomes the next project's starter, so you cannot easily skip ahead.

Requirements and time. You need an agent that can edit files and run commands, such as Claude Code or Codex. Without one you can read the material but not complete the projects. The stated pace is about one week per phase part-time, with phases 1–3 doable in a long weekend.

Limits

• Inconsistent counts. The README says "13 lectures / 7 projects" and "all six course projects" in places, but 14 lectures and 8 projects in the outline and repo tree. Recently added material may not be reflected everywhere.
• No licence stated on the fetched page. Check the repo before reusing the templates or code.
• No measured results of its own. The course says projects "compare weak vs. strong harness results," but this page reports none.
• Not aimed at prompt-only or no-code users, by the course's own statement.