Agent Engineering Handbook

Build better coding environments for agents: turn a repeated correction into an executable check, a confusing decision into discoverable knowledge, and a missing operation into a usable tool.

An independent, public handbook maintained by Desmond Landry (@desland01), based on Theo's video (Claude Code's creator has some really good advice, July 21, 2026) and extended through direct inspection of Theo's T3 Code and Melee fork, Matt Pocock's Course Video Manager, and Boris Cherny's public engineering work. With attribution and thanks to Theo (@t3dotgg), Matt Pocock (@mattpocock), and Boris Cherny (@bcherny) — none of whom endorses this handbook. Research snapshot: September 9, 2026; later pull requests are identified as later evidence.

The source-derived facts (what the video says, what the inspected repositories contain) are kept separate from author recommendations and from this edition's own local checks. See validation.md for exactly what was run and what was not.

Start with the problem you have

Read the live handbook or its GitHub source. The map is the shorter entry point: every guide, investigation, skill, idea and frame on one page, grouped by source. The screenshot gallery contains 12 ffmpeg frames with timestamps, observations and full-size images. A post visible on screen is distinguished from a live demonstration; several examples are narrated only.

All 13 implementation guides

1. 01 — Convert recurring failures into permanent rules
2. 02 — One critical journey test beats broad shallow coverage
3. 03 — Make previews usable from the agent's environment
4. 04 — Let the agent read and resolve CI failures
5. 05 — Put project knowledge where the next task needs it
6. 06 — Build small tool adapters for capability gaps
7. 07 — Turn newcomer questions into improvements
8. 08 — Keep contracts consistent across the application
9. 09 — Verification contracts
10. 10 — Codebase navigation and tooling
11. 11 — One domain language and one transport when an agent calls your API
12. 12 — Make pipeline artifacts reusable, resumable and honestly complete
13. 13 — Validate the output through several useful views

Each guide gives a concrete method, fitting use cases and verification limits. Code sketches are labeled as illustrative; the bundled lint example is separately runnable and tested (see examples/recurring-rule for the npm demo).

What the repository investigations add

• Theo: T3 Code and Melee — architecture, custom lint, native skill dispatch, preview evidence, all 13 fork PRs, the linked verifier commit, and the source-completeness verifier repair.
• Matt Pocock: Course Video Manager — domain vocabulary, one HTTP transport, static import boundaries, publication identity, resumable agent work, subprocess output draining and gaps between documentation and CI.
• Boris Cherny: public engineering work — a staged compiler architecture, seeded fuzzing, generated-output validation, benchmarks, CLI compatibility, narrow tools and instruction-validator repairs.

The reports distinguish inspected code, author-reported results, remote CI records and local execution. They link to pinned source snapshots wherever possible. Finding code under an account does not mean the account owner personally authored every change; open PRs are separated from merged work.

Four portable skills

Standalone, portable skill directories you can adopt into any agent setup that reads Markdown instructions. Each keeps its reusable expertise and its owner-authorization boundaries, and ships scoped references plus OpenAI-compatible interface metadata.

Each directory includes SKILL.md, references/implementation.md, references/source-patterns.md, and agents/openai.yaml interface metadata whose default prompt references the skill's own $skill-name. See adoption.md for how to install one or combine it into an existing skill. Do not load all four for every task.

Use the relevant guide with your existing workflow today. Task prompts make the input and expected evidence concrete.

The 19 useful video ideas, in order

These are distinct takeaways, including qualified opinions. The table maps each to implementation guidance. See detailed extraction and structured ideas for speaker attribution, evidence type and caveats.

What was checked

The local lint demonstration proves that the unwanted import passes without its restriction and is rejected with the restriction while the supported import passes. The Melee verifier's original nine tests and current fifteen tests pass; a separate synthetic comparison reproduces the completeness gap and the repair. These are verifier tests with mocked builds, not a real game build.

See validation.md for the exact scope, screenshot hashes and rendered-page checks. The public applications, upload service and private agent systems were not executed. Sponsor claims and universal career/productivity claims remain unverified.

Build and preview locally

The site is static HTML generated from the Markdown in this repository; the generated public/ directory is committed for local reading. Vercel rebuilds it from the editable sources and checks the result on each deployment.

# Regenerate public/ from the editable sources (idempotent):
python3 -m pip install -r build/requirements.txt   # Markdown==3.10.3
python3 build/render.py
python3 build/check.py                             # links, counts, hashes

# Preview locally (any static file server works):
cd public && python3 -m http.server 8000
# open http://localhost:8000

build/render.py reads only the editable Markdown, assets and JSON — never public/ — and rewrites public/ from scratch, so rerunning cannot feed on its own output.

Run the lint demo:

cd examples/recurring-rule
npm ci --ignore-scripts
npm run demo        # detects the bad import; accepts the supported import

Keep or rebuild this handbook

The Markdown, images, JSON evidence and four skill directories are the editable sources; everything in public/ is generated. build/render.py regenerates the pages and gallery using Python Markdown 3.10.3 (build/requirements.txt pins that dependency). See CONTRIBUTING.md for the update flow and ATTRIBUTION.md for source attribution. The original long-form video is not redistributed here and no complete transcript or repository clone is bundled.