This repository is a retrieval-optimized reference bundle for the 12 theses of harness engineering. It follows the architectural principles it teaches:
- Route context just in time — The root AGENTS.md routes to specific theses. Don't preload the entire corpus.
- One concept, one owner — Each thesis lives in exactly one file under
theses/. - Progressive disclosure — Each thesis starts with a one-paragraph summary, develops the argument, links to primary sources for deeper reading, and ends with an actionable checklist.
| Path | Owns |
|---|---|
README.md |
Public definition, installation, and quick start |
AGENTS.md |
Application routing and thesis index for consuming agents |
ARCHITECTURE.md |
Document ownership and routing invariants |
cli.js |
CLI entry point — list, scaffold, and thesis-reading commands |
theses/ |
One file per thesis, numbered and self-contained |
theses/_scaffold/ |
Template files for harness-kit scaffold |
- One navigation contract. AGENTS.md is the root map. It names each thesis, its file location, and the decision it addresses.
- Route just in time. A normal task reads only the theses relevant to the unresolved decision. Primary sources load only when they can change a decision.
- Separate argument from evidence. Each thesis synthesizes the argument and links to primary sources (lopopolo/harness-engineering, hyperbo.la, Twitter, talks). The kit does not duplicate evidence it cannot maintain.
- Keep the bundle proportional. This is a reference, not the full anthology. File count and depth are deliberately bounded for quick retrieval.
- Test behavioral effect. The kit succeeds when it changes engineering decisions, not when someone repeats its terminology.