Close the CI feedback loop: let the agent trigger CI and read failed logs itself
Sponsor segment (Blacksmith), read by Theo
Any repo where agents open PRs and CI results gate them, especially when humans currently relay errors by hand.
Build a codebase agents can work in
Turn recurring agent mistakes into checks, clear instructions and usable tools. This handbook connects nineteen video ideas to thirteen guides, four portable skills and three source investigations. Each guide states what its checks can prove.
Edition of September 9, 2026. An independent public edition by Desmond Landry (@desland01), built from Theo's video and direct inspection of three repositories. Theo, Matt Pocock and Boris Cherny do not endorse it. Full attribution is at the end of the page.
Start with the problem you have
Each row names a situation and the guide that addresses it.
- You keep correcting the same mistake01 Recurring failuresRunnable lint example
- Tests pass but the actual feature fails02 Critical journey tests09 Verification contracts
- A new agent cannot run or inspect the app03 Preview workspaces
- You relay CI errors by hand04 CI feedback
- Agents repeatedly miss project decisions05 Knowledge and instructions
- The agent cannot perform a required operation06 Tool adapters
- Your project is becoming hard to navigate11 Domain language and APIs10 Codebase navigation
- Long jobs repeat expensive work or accept bad output12 Artifact identity and recovery
Start with the advice behind the methods
Browse the ideas in video order, or open the full set grouped by practice, advice and opinion.
Design tiny two-actor end-to-end tests that catch whole classes of failure
Theo (recounting his time at Twitch)
Distributed or multi-party features (chat, notifications, sync, webhooks) where component tests pass but integration breaks.
Infra and DX automation now speeds up every agent, not just you
Boris (as quoted by Theo), endorsed by Theo
When deciding where to invest tooling effort on a team running multiple concurrent agent sessions.
Preview environments matter more now that code is built outside your machine
Theo
Any agent workflow where the human cannot or should not run the app locally to review each change.
Build small tool adapters for things the CLI cannot do (video/asset uploads to PRs)
Theo
Recurring agent tasks blocked by a missing API surface: media attachments, screenshots, signed artifacts, anything possible by hand but not by tooling.
Authoring and testing a small skill has a tight, low-stakes feedback loop
Theo
Individual contributors hesitating to invest in tooling because it used to be high-effort, low-recognition work.
Teams are now more willing to fund tooling time — use it
Theo
Engineers who assume management still treats deep tooling as slacking.
Move recurring fixes from per-occurrence agent corrections into executable checks (lint rule, CI step, routine)
Boris (as quoted by Theo)
Any recurring correction — style, safety, architecture — that is currently handled by repeated review comments or repeated agent fixes.
Custom lint rules became economical: agents lower the cost of the code and its tests
Theo
Teams with a backlog of issues always caught by eye that were never worth the engineering cost.
Encode domain knowledge as infrastructure so newcomers' agents get steered too
Boris (as quoted by Theo), partially endorsed by Theo
Codebases where the same contextual rules are re-taught verbally to every new engineer.
Treat newcomer questions as a signal for onboarding gaps (Theo's dumb-questions practice)
Theo
Any team onboarding engineers, and equally for agents: an agent's repeated confusion is the same signal.
Write your CLAUDE.md / AGENTS.md yourself; watch agent behavior and adjust
Theo
Anyone whose agent instructions were generated once and never deliberately revised.
Use steering files to make the agent say no (and to get fast feedback when things go wrong)
Theo
Projects that attract repeated out-of-scope requests from users, teammates, or agents.
Compose layers so type safety runs end to end — and expect that click more often now
Theo
Any stack where hand-written duplicates of a single source of truth drift apart.
Write Claude.mds, review.mds, skills, and docs so agents work with zero prompting context
Boris (as quoted by Theo), endorsed by Theo
Any shared codebase; the test is whether a stranger's first prompt succeeds.
Calibrate from a cold start: run minimal prompts first, add context only where it fails
Theo
New projects, newly agent-enabled repos, and any team tempted by large plugin lists.
Instruction files should steer toward success, not list where things are
Theo
Auditing any existing instruction file that reads like a directory listing.
Building environments where code lands well is a career-level skill (with a grain of salt)
Theo (self-flagged as speculative)
ICs aiming at senior/staff scope, and teams deciding what to recognize and promote.
Solo projects now exceed your own comprehension — build systems so you and your agents don't get lost
Theo
Solo developers and small projects where all context is tacit and agent-generated code outpaces personal recall.
Choose a skill for the problem you have
Each skill is a folder of instructions for an agent setup that reads Markdown. Choose the one you need rather than loading all four for every task.
- Agent feedback engineering
Choose and prove a reusable response to an observed failure.
4 ideas feed it
Open the skill skills/agent-feedback-engineering/ - Agent-ready workspaces
Repair the setup, preview or outcome-check gap that blocks a fresh agent.
8 ideas feed it
Open the skill skills/agent-ready-workspaces/ - Agent context calibration
Diagnose missing project knowledge and select the right place for it.
4 ideas feed it
Open the skill skills/agent-context-calibration/ - Agent tool adapters
Bridge a missing capability and verify the intended agent can use it.
3 ideas feed it
Open the skill skills/agent-tool-adapters/
Find a guide for your next change
Thirteen guides grouped by source, each with a method, fitting use cases and verification limits.
Check the sources behind the guides
The three reports separate inspected code, author reports, remote CI records and local runs. They link to pinned snapshots where possible and distinguish open pull requests from merged work.
See what was on screen
Twelve frames extracted with ffmpeg separate visible evidence from narration. Each opens the gallery for full-size images, observations and extraction hashes.
What was checked, and what was not
- 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.
Attribution
- 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 ATTRIBUTION.md.