Fix the environment and the agent works
You have corrected the same agent mistake three times this week. You fix it by adding a check, a working preview, or one sentence to the right file. Ten short lessons show you where those fixes live, and each ends with one change you can make today.
Start with the problem you have
You will recognise at least one of these ten problems from your own week. Open its lesson and you leave with one fix you can try today.
- You keep correcting the same mistakeLesson 1 You turn repeated mistakes into reliable checks
- You relay CI errors by handLesson 2 You diagnose failed checks without copying logs
- Tests pass but the actual feature failsLesson 3 You test the result users actually need
- A new agent cannot run or inspect the appLesson 4 You give agents a working preview
- The agent cannot perform a required operationLesson 5 You make missing operations usable by agents
- Agents repeatedly miss project decisionsLesson 6 You write instructions that fix observed confusion
- Types, docs and runtime disagreeLesson 7 You keep shared contracts consistent across layers
- Your project is becoming hard to navigateLesson 8 You navigate unfamiliar code without guessing
- Long jobs redo finished workLesson 9 You resume work without repeating completed steps
- Every new contributor hits the same setup wallLesson 10 You remove obstacles for the next contributor
Name what keeps going wrong
Start with the failure you saw this week, in your own words, before you look for a lesson.
Read the one lesson that covers it
Each lesson takes a few minutes and names the sources behind every claim it makes.
Try one change and check it today
Every lesson ends with one concrete action you can make and then prove worked.
Ten lessons each end in one change
Read them in order, or jump to the problem that cost you an afternoon this week. Every lesson names the sources it draws on and ends with one action you can try today.
- Lessons 01–03Stop repeat workChecks, CI access and real verification instead of repeated corrections.
- Lessons 04–06Give agents what they needPreviews, tools, instruction files and cold-start context.
- Lessons 07–10Keep work understandable and recoverableShared contracts, navigation, recovery and contributor improvements.
Seven skills carry the lessons into your agent
Your agent reads a skill file before it starts, and you pick the one that fits today's task. Read the explanation first, and treat the included evaluations as unrun.
- Agent feedback engineeringChoose and prove a reusable response to an observed failure.
- Agent-ready workspacesRepair the setup, preview or outcome-check gap that blocks a fresh agent.
- Agent context calibrationDiagnose missing project knowledge and select the right place for it.
- Agent tool adaptersBridge a missing capability and verify the intended agent can use it.
- Agent output verificationJudge whether a run's actual output satisfies what was asked.
- Agent artifact recoveryResume interrupted work without repeating completed stages.
- Agent contract consistencyKeep vocabulary, shared types and runtime behavior in agreement.
See where every claim comes from
Nothing here rests on a post you have to take on trust. Three inspections separate what we read in the code, what the authors reported, and what we ran ourselves.
- Report 01Theo: T3 Code and MeleeThe inspection separates advice, inspected code and measured results.
- Report 02Matt Pocock: Course Video ManagerThe Course Video Manager inspection: shared vocabulary, one transport, durable artifacts.
- Report 03Boris Cherny: public engineering workThe inspection compares several ways to check useful output.
Every check below names what it does not prove
- RUNThe lint demo accepts the unwanted import without its rule, rejects it with the rule and accepts the supported import.
- RUNThe Melee verifier passed its original nine tests and the current set of fifteen.
- RUNA synthetic comparison reproduced the completeness gap: the original verifier accepted incomplete source metrics, while the repaired version rejected them.
- LIMITThe verifier tests used mocked builds and establish verifier behavior only. No real game build was run.
- NOT RUNThe public applications, upload service and private agent systems were not executed for this research.
- UNVERIFIEDSponsor performance claims and universal claims about careers and productivity have not been verified.
Credits identify the sources and limits
- Author
- Desmond Landry @desland01
- Sources
- Theo, Matt Pocock, Boris Cherny @t3dotgg · @mattpocock · @bcherny
- Endorsement
- Theo, Matt Pocock and Boris Cherny do not endorse this handbook.
- Snapshot
- September 9, 2026 later pull requests are identified as later evidence
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.
Frames are short excerpts from the video, reproduced for identification and commentary; they remain © Theo / their original owners. Full attribution in Credits and reuse.