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.
What backs Theo's advice?
Explore nineteen ideas from Theo's July 21, 2026 video, grouped by practice, advice and opinion. Each idea links to the moment in the video where it was said.
What he actually does
Theo describes practices he uses and recounts his experience of work in real codebases.
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.
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.
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.
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.
What he tells you to do
Theo gives his own advice and quotes Boris, with endorsement or qualifications where recorded.
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.
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.
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.
Write your CLAUDE.md / AGENTS.md yourself; watch agent behavior and adjust
Theo
Anyone whose agent instructions were generated once and never deliberately revised.
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.
What he offers as opinion
These ideas are offered as opinion, including several the video explicitly qualifies as such. This group also includes the sponsor segment, whose performance claims remain unverified.
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.
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.
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.
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.
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.