Skills library
129 skills from 55 videos and 11 T3 Code releases. Seven are new in the last 7 days.
Have an AI model build 3D models, scenes or even simple games in Blender, the free 3D modeling app, starting from a screenshot or a description. The model controls Blender for you from the command line.
In 2 videos, first seen Sep 4Run your main agent on a long goal, and keep a second chat that reads its history. You use that second chat to ask for status and what is slowing the work, which surfaces the real blockers.
In 1 video, first seen Oct 28T3 Code can run coding agents from the official ACP Registry or any installed ACP command. The agent brings its own models, tools, and sign-in. T3 Code provides projects, threads, checkpoints, and task delegation.
From the T3 Code guide. In 1 T3 Code release, first seen Oct 15Have a coding agent rebuild an app on another platform, using the original app as its reference. It gets the original source code, a simulator (a virtual phone on your computer) and the working app to copy. When something is off, you point it at how the original behaves.
In 2 videos, first seen Sep 6Instead of browsing GitHub, ask an agent to list recent pull requests (proposed code changes) in the areas you care about, each with a short summary. It replaces slow, tab-heavy GitHub browsing.
In 1 video, first seen Sep 15Lets your coding agents work with pull requests for you. They can read, open, update, respond to and merge PRs through MCP tools, which are plug-in tools an agent can call.
In 1 T3 Code release, first seen Nov 20Paste a large pull request (a proposed code change) into Claude and get a plain breakdown of what is changing, why, and where the risk is. You can also have Claude build a visual artifact that teaches how that part of the code works.
In 1 video, first seen Nov 6Write your AGENTS.md (or CLAUDE.md), the instruction file every agent reads at the start, around what your product is, what makes it special, and a glossary of its terms. This gives the agent values to protect and shared words for talking with you.
In 1 video, first seen Aug 25A way to write a post about a price or limit change. Lead with the bad news, state the real net effect in plain numbers, and acknowledge how users will feel it instead of spinning it as good.
In 1 video, first seen Aug 30Turn on Anthropic's design skill, a short instruction file the model loads, to get better-looking web pages and interfaces. It steers models away from generic, cluttered designs toward pages with real ideas, unique looks and nice animations.
In 3 videos, first seen Oct 8A line from Anthropic's guide: if the agent finds bugs or performance issues the task didn't mention, it should not fix them in this change, but list them as follow-ups in its summary. Anthropic says this cuts unrequested additions sharply with no drop in task success.
In 1 video, first seen Sep 22Auto mode lets Antigravity's own review approve routine steps, so it only stops to ask you when it actually needs you.
In 1 T3 Code release, first seen Nov 20Before reviewing a big agent-made change, ask the agent how risky it is to merge today and what the worst outcome would be. The answer tells you whether reading the whole change line by line is worth it.
In 1 video, first seen Sep 24Ask the agent in your instructions to tell you what it's doing during long runs. Some models, like Fable 5.1, otherwise work silently, for example making 30+ tool calls in a row with no text.
In 1 video, first seen Sep 3Ask one agent to describe in detail what it can do, then hand that description to another agent and have it copy those abilities. Theo did this to bring Codex's computer-use abilities into T3 Code using Opus.
In 1 video, first seen Oct 18Astra in OpenAI's Codex app can operate your Mac directly: clicking, opening apps and doing tasks in the background. It suits prep work, like setting up Final Cut before video editing.
In 1 video, first seen Sep 10Astra can launch many helper agents (sub-agents) that message each other and report back, and it uses their updates to keep the group on track. Fable, by contrast, plans its sub-agents in fixed steps up front.
In 1 video, first seen Sep 10Claude Code can store 'memories' about your projects. Theo had his agent list them, then disable the feature across all his machines, archive the existing memories and delete them.
In 1 video, first seen Aug 25Have the model audit your codebase and propose improvements, then tell it to spin up sub-agents (parallel helper agents) to carry them out as separate pull requests. This gets many fixes done at once.
In 2 videos, first seen Sep 3Write a 'babysit' skill, a short saved instruction file, that tells the agent to watch its pull request. It fixes failing checks, updates the change onto the latest code, handles review-bot comments, and keeps going until everything is green.
In 3 videos, first seen Sep 4BAML is a small language for declaring the exact shape of answer you want back from an AI model. Its runtime repairs badly formatted JSON, so your app always gets data in the promised shape.
In 1 video, first seen Sep 20Put screenshots of rival products into the review panel, so reviewers score your app side by side with competitors instead of on its own. The goal becomes beating those apps, not just your own earlier version.
In 1 video, first seen Nov 3A setting in Claude Code that caps Haiku 5.5's auto-compact window at 100k. The session then summarizes itself before it passes the price threshold. It applies only to Haiku, sub-agents included, so Haiku sub-agents never get expensive.
In 1 video, first seen Oct 8Have a model label a large dataset and give each label a confidence score from 0 to 1. A slider sets how sure the model must be before a label counts. Theo used Jev on 32,311 messages from 1,118 of his own agent chats.
In 1 video, first seen Sep 20Anthropic gives Claude Slack channels, each focused obsessively on the health of one system component, with access to its logs and metrics.
In 1 video, first seen Oct 15Use Haiku, Anthropic's smallest and fastest model, to sweep large amounts of data or every file in a codebase for a pattern that ordinary code can't detect.
In 1 video, first seen Nov 12Use Anthropic's Opus 5.5 for longer, multi-step agent work. Its strength is reliability: it is less likely to do something stupid, so you can trust it to stay on track for longer.
In 1 video, first seen Sep 27CLI Proxy API is a small server you sign into with your Claude and Codex accounts. It keeps those logins alive and gives Claude Code and Codex one address to send requests to, so you can juggle several accounts from one place.
In 2 videos, first seen Oct 1You can use a Codex subscription on CI, the automated system that checks code changes, so it reviews each pull request (a proposed code change). OpenAI says this is 'mostly allowed' as long as it doesn't serve user traffic or charge others for access.
In 1 video, first seen Oct 7OpenAI's Codex coding subscription. Its $100 plans currently have no 5-hour limit, only a weekly limit, which makes it an alternative to Claude Code plans.
In 1 video, first seen Aug 30Context compaction shrinks the earlier parts of a long conversation so the agent can keep working, and it works across the different agent harnesses T3 Code supports. A redesigned usage insights view lets you see how much you are using.
In 2 T3 Code releases, first seen Aug 25Connect your AI model to Gmail and Notion and ask it what you should be prioritizing. It reads your email and notes and lists what needs attention, including things you have dropped.
In 1 video, first seen Sep 4ChatGPT lets you link several email accounts at once. You can then ask it about your inbox instead of reading everything yourself.
In 1 video, first seen Oct 18Council is a skill (saved instructions for an agent) where your coding agent asks reviewers from several model families to check a change, plan or design and score it from 1 to 10. You set a minimum score that depends on the work, and every reviewer must reach it before the change is done.
In 4 videos, first seen Nov 2One model writes the code and a model from a different lab reviews it before you merge. The reviewer brings a different view, like a friend at another company, and catches problems the writer missed. They can go back and forth until both agree.
In 7 videos, first seen Sep 22Make merging safe on two fronts: code is very likely fine by the time a person approves it, and any mistake that slips through goes to a few users first and is easy to revert (undo).
In 1 video, first seen Oct 15When an agent hands research to sub-agents (helper agents it launches), tell them to read the actual code behind every claim and never cite a file they didn't open. They should also not edit anything. Muse does this on its own.
In 1 video, first seen Oct 13Have agents attach a video or screenshot showing their change working, then watch it yourself. A GitHub CLI update lets agents upload these without adding them to the codebase.
In 1 video, first seen Aug 24Dictation lets you speak to T3 Code instead of typing. You can turn it on for this device.
In 1 T3 Code release, first seen Nov 9Remove the 'anti-slop' writing-style rules from your agent instruction file (the CLAUDE.md the agent reads every session). Then see how the model writes on its own.
In 1 video, first seen Sep 23First build an eval (an automated test set) from real cases where agents misused your product. Get to a perfect score, then shrink prompts or tool descriptions while keeping that score. Theo cut T3 Code's tool descriptions to a third of their size this way.
In 1 video, first seen Nov 3Expo is a toolkit for React Native apps. One developer sets up a development version, others scan a QR code and start changing it, and updates to users get easier. Its simulator and web support let agents test changes without a native build.
In 1 video, first seen Sep 14Using the Fable 5.2 model to review batches of pull requests (PRs), the code changes waiting to be merged. When Fable says a PR is good to go, it usually is. It skips minor nitpicks, though it misses some things.
In 1 video, first seen Nov 10Give an agent a tool that looks real, like 'run a terminal command', but never runs anything. A small, cheap model invents believable output, so you can watch what the agent tries with nothing at risk.
In 1 video, first seen Sep 1Start loading a page's data when the user hovers over a link, before they click. Anthropic did this for conversations. It makes the app feel faster for a small change.
In 1 video, first seen Oct 5When a model keeps doing something you dislike, ask it what the pattern is called and why it does it, then add a rule against it to your agent instructions file (AGENTS.md).
In 1 video, first seen Nov 10Run separate agent threads that watch your working agents, find blockers and repeated failures, and change their setup so many agents can run in parallel without colliding.
In 1 video, first seen Oct 10Keep one repo (a project folder) that documents every computer you use: how to connect, what it's for, what is installed. Agents work from it to do machine chores and set up new machines from a single prompt.
In 2 videos, first seen Sep 15You can fork any thread in T3 Code, which makes a copy you can take in a new direction while the original stays as it was.
In 1 T3 Code release, first seen Oct 15Ask the AI to copy (fork) an existing tool you use and change it to fit how you work, instead of only building small new apps.
In 1 video, first seen Oct 20Full Send is a skill that tells the agent you aren't watching. It finishes the change, proves it works, opens the pull request, handles reviews and failing checks, and merges on its own, only bothering you if it gets stuck.
In 1 video, first seen Oct 26Full Stop is a version of Full Send where the agent keeps fixing, verifying and handling reviews until the change is ready, but never merges. You make the final call.
In 1 video, first seen Oct 26A low-cost model from GLM: 15 cents per million input tokens and 50 cents per million output, often half off. It is good at long agentic tasks, and Theo says it matches or beats GPT-5 High on benchmarks.
In 2 videos, first seen Oct 11You can keep using OpenAI's GPT models inside the Cursor editor after the cutoff. Use your own OpenAI API key, a password-like code that bills your own OpenAI account, or OpenAI's Codex extension inside Cursor. OpenAI says both routes will keep working.
In 1 video, first seen Aug 30GPT-6.1 Sol is an OpenAI model you can use for analysis and review rather than writing code, such as explaining a large codebase or auditing for bugs.
In 2 videos, first seen Oct 14Give Muse a bug report and ask it to explain the cause and check whether someone already has a fix open. A pull request (PR) is a proposed code change waiting to be merged. On a real T3 Code bug, Muse explained the cause readably and found a smaller, cleaner fix someone had already filed.
In 1 video, first seen Oct 13After a long Astra run, switch the same thread to Claude Opus 5.5 and have it explain what was done. This gives you a readable account of the changes and the reasons for them.
In 1 video, first seen Nov 18Hand one model's work to a stronger model and ask for an honest check. A pull request is a proposed code change. The second model can find where the work went wrong, locate the real cause and propose or write a better fix.
In 2 videos, first seen Oct 14When a post shares cost figures that seem off, screenshot it and ask Claude or Codex to check the math and assumptions.
In 1 video, first seen Oct 26When your systems show failures that shouldn't happen, have Claude add alerting, monitoring and tracing. Tracing records where and when the failures occur and in which regions. Claude then writes up the findings. Theo's report found serious flaws in Cloudflare's pipeline, which Cloudflare said it would fix.
In 1 video, first seen Oct 30Keep your running app and a fast model side by side. Send small change requests one at a time and watch each one land. This works well for interactive parts of a product, such as buttons and flows. You refine through many quick rounds instead of hoping for a perfect first try.
In 2 videos, first seen Oct 6The Issues page shows issues from Linear, GitHub, GitLab, Azure DevOps and Forgejo next to your threads. You can filter by label and assignee, sort, save views, link issues to threads, and hand an issue to an agent all the way to a merged PR.
In 1 T3 Code release, first seen Nov 20Use Jev as a quick first pass for high-volume, simple sorting jobs, like clearing low-effort spam from email, before an expensive model touches them. In a clip, it averaged about 200 ms per email, 38 emails per second. It suits jobs too cheap or too frequent for an LLM.
In 1 video, first seen Sep 20just-bash is a library by Malte at Vercel that gives agents a fake bash command line in memory, with a fake file system and processes. Agents can explore files as if on a real computer without a real machine. Alone it can't build a real app, since you still need a TypeScript compiler; TS-Rust fills that gap.
In 1 video, first seen Oct 10Kiro can be chosen as the provider, meaning the service that runs the agent, inside T3 Code. It is in beta as an Early Access provider.
In 1 T3 Code release, first seen Nov 4Legend List is a list library by developer Jay for long scrolling lists. It draws only the rows on screen. It is written entirely in JavaScript, so it also runs on the web. It's why T3 Code's scrolling is smooth in its desktop and web apps.
In 1 video, first seen Sep 14Let each agent launch its own copy of your app and record a video of the change working. It can use fake, pre-recorded model responses so full end-to-end tests don't spend AI tokens. By the time the agent says code is ready, it has already checked the change most of the ways Theo would.
In 1 video, first seen Nov 6Record every AI request and response in a database you control, so you can check your real costs and how models behave. Theo routes all his usage through his own proxy, a middleman that relays requests to the providers. It stores everything in ClickHouse, a database built for huge volumes of data, on his home storage server.
In 2 videos, first seen Oct 26Run agents autonomously overnight, not mainly to get code written but to hit failure cases. A snag agents hit once every two hours shows up two or three times in a five-hour run. Fixing those spots makes shorter, supervised runs go smoothly.
In 1 video, first seen Sep 17Give an agent a standing goal with a measurable target and tell it to keep going until it hits that target. This keeps it working instead of stopping after one fix. Theo's target of 'minimum 2x faster than the Go build' kept his agent working for two weeks until it got there.
In 1 video, first seen Oct 10Luna is OpenAI's small, fast, cheap 5.6 model for background chores like titles, status updates, summaries and sorting threads. It can call tools and turn messy context into a JSON object, which is structured data your tools can read. It costs a little under half of DeepSeek's V4 Flash per task.
In 1 video, first seen Aug 22Give the agent standing instructions for long runs to check its own work before reporting back. It tests changes through computer use, addresses the AI code review bots' comments and has sub-agents do another pass. It then returns only once it is fairly confident nothing users would notice has broken.
In 1 video, first seen Sep 15Set up your project so anyone, human or agent, can quickly start the app and try a change. Theo made T3 Code's dev server (the local test version) run over Tailscale, a private network, with a read-only snapshot of real data. That way he can check agents' changes and agents can check their own.
In 1 video, first seen Aug 24Pick models by what a whole task costs, not by price per token. For bulk reading, input price matters most. For agent coding, caching (cheap reuse of already-read text), reliability and token efficiency matter more. You can also give a specialist job, such as design, deep review or game assets, to a model that is stronger at it.
In 4 videos, first seen Sep 23MCP (Model Context Protocol) is a standard connection that lets an agent reach data and services outside its own tool. For example, it can query your product database on Supabase or answer questions from your PostHog analytics.
In 1 video, first seen Oct 30Have several models each review the same codebase and suggest areas to improve, then use strong models as judges to score the suggestions. In Theo's test, Grok 4.7 came just behind Astra on viability, and Fable 5.1 made fewer suggestions that still scored well.
In 1 video, first seen Sep 22Have the agent run every task on two models, then have a panel of judge models score results without knowing which model did which. The models can also review each other's plans. It showed Theo that Opus did more correct, deeper work while Muse did more work, faster.
In 1 video, first seen Oct 13A terminal command that gives Claude Code or Codex the context they need to debug T3 Code on your own machine. It can also file an issue with the T3 Code team when the problem is not just your setup.
In 1 video, first seen Oct 29Open-source apps that put Claude Code, Codex and other agents in one window while you pay each AI lab directly. Theo names his own T3 Code, plus Gene (from the Coolify developer), Superset and Herder.
In 1 video, first seen Aug 30From one thread, ask the agent to start helper agents on different models that work on the same task. T3 Code's delegate_task lets an agent start child agents on any provider or model and wait for their results. Each helper runs in its official tool on a subscription you already have.
In 1 video, first seen Oct 15An automatic check that runs on every pull request (a proposed code change) and measures one thing you care about, such as how much data a page sends. It posts the number on the change and flags or fails it when that number gets worse, so the gains you made stay in place.
In 3 videos, first seen Aug 25Effort level controls how long the model thinks and how much it checks before answering. Picking the right level lets you match the model's thinking to the task. Opus offers low, medium, high, extra-high and max.
In 2 videos, first seen Sep 23pnpm, a package installer, can use hard links: many file names point to one stored copy instead of duplicating files. The result depends on the file system underneath. On Linux XFS one reinstall test fell from 7 to 2.85 seconds; on a Mac installs grew from about 40 to about 100 seconds.
In 1 video, first seen Aug 20A skill (saved instructions for an agent) that hands a fresh agent an existing pull request to finish. The agent reviews it, decides whether to build on it or only use it as a reference, then pushes and manages the branch until it lands, crediting the original author.
In 2 videos, first seen Sep 3Have a cheap, fast model go through all your open pull requests and sort them, for example by ease of merge and value to users, or into merge, close or needs another pass. This tells you quickly what to act on first, for little money.
In 3 videos, first seen Aug 22For 3D or interactive projects, Astra builds the first version to get the look roughly right. Then Fable does the detail work on animation curves, camera movement and overall feel.
In 1 video, first seen Sep 10Delete automated tests that add no value. When a pointless test shows up again, ask the agent that wrote it why, and fix that root cause instead of only deleting the test.
In 1 video, first seen Nov 9T3 Code has a pull requests page that works across providers, where you can review pull requests (proposed code changes) in the app. You can send requests about specific lines to the agent, scan comments more easily, take a guided beta tour of a pull request, and @mention users when writing.
In 4 T3 Code releases, first seen Aug 25When an agent's code is wrong, change the prompt or context and rerun instead of fixing the output yourself. The idea is that some input produces code needing no edits, and once you find it you can repeat the task without new effort.
In 1 video, first seen Nov 19Keep a set of prompts that earlier models failed at. When a new frontier model comes out, run that set before your usual prompts. It shows what the new model can now do.
In 1 video, first seen Sep 17Have agents build a narrow, custom version of a paid tool that fits only your needs. Theo had agents copy all his analytics data from a provider into his own ClickHouse database (a fast analytics store) on R2 (Cloudflare file storage).
In 1 video, first seen Oct 15When Astra loops on a pull request (a proposed code change) without finishing, stop it and give the PR to Fable 5.1 to finish and land cleanly.
In 1 video, first seen Sep 10Rewinding lets you take a conversation back to an earlier point while keeping the file changes already made.
In 2 T3 Code releases, first seen Sep 15A Mac tool that makes git work trees (separate copies of your code for agents) much faster. It shows the agent a full copy of the files but only writes new files when they change.
In 1 video, first seen Aug 20T3 Code can run scheduled tasks. Each task links to its thread, so you can open the conversation where the work happened.
In 2 T3 Code releases, first seen Oct 15Have each review agent score a change, for example from 1 to 10, and set the score a PR must reach, instead of looping until there are no comments. You can also ask for scores on what you care about, such as how likely you are to be happy with it.
In 2 videos, first seen Nov 6A way to make a fuzzy goal like "looks as good as the real game" checkable. After each good build, the agent screenshots set moments, compares them with reference footage, and a panel of AI judges says whether it looks good enough.
In 1 video, first seen Oct 28Paste a screenshot of a user's bug into your agent and have it fix the bug, test the fix, record a video proving it works, open a PR (a proposed code change) and handle review comments. The agent carries the whole job from screenshot to working PR.
In 1 video, first seen Sep 17Use your agent instructions file (AGENTS.md, standing instructions the agent reads every session) to set how it talks to you. Theo's asks for short, easy-to-read answers, varied sentence length and simplified technical English, and mentions he has ADHD.
In 1 video, first seen Oct 13Set up the machines your agents work on around what the agents handle well, not your own preferences. Theo is building T3OS, a customized Ubuntu image that holds what his agents need.
In 1 video, first seen Oct 20You can share a thread with a guest who watches the agent work live. The guest can send messages, and they only reach the thread if you approve them.
In 1 T3 Code release, first seen Oct 27A side chat lets you ask a question about a thread without adding to it. It starts with what the thread knows, and the thread keeps working while you ask.
From the T3 Code guide. In 2 T3 Code releases, first seen Oct 31Size hardware for running an AI model on your own machine by checking that the whole model fits in the graphics chip's memory. Memory amount, not chip speed, decides what you can run.
In 1 video, first seen Aug 28Keep a smart model like Opus as the main agent and let it hand small jobs to cheaper, faster sub-agents (helper agents it launches itself), such as Haiku 5.5 or Sonnet 5.5. The main model plans and directs; the helpers do the many small attempts or investigations in parallel.
In 2 videos, first seen Sep 29Write a detailed bug report for your agent saying what is wrong and what the fix should look like, and give it even a minimal way to check its fix. This lets the agent solve the bug reliably.
In 1 video, first seen Aug 24Give each AI provider's subscription a specific job instead of using one for everything. Theo uses OpenAI for computer use and verification, Claude for day-to-day coding, and Grok only for its Grok bot assistant.
In 1 video, first seen Nov 2One command hands a written support playbook to Claude Code or Codex. The agent asks what went wrong, gathers reproduction steps, checks the code, then fixes the problem or files a clean bug report with secrets removed.
In 1 video, first seen Oct 18T3 Code is Theo's free, open-source app for running many coding agent threads. It sits on top of harnesses like Claude Code and Codex, using your existing subscription, skills and plugins. You can manage threads like an email inbox, run each on a machine you pick, and let your agent start threads on another harness such as Codex.
In 3 videos, first seen Oct 1T3 Connect lets you run agent threads on a spare computer, such as an old Linux box, and manage them from the T3 Code desktop or phone app. You install it with one command. It is free for three devices.
In 1 video, first seen Oct 29A terminal command in T3 Code, t3 thread, that lets you start a thread, send it a message, and read it without leaving the terminal. Handy for checking on or nudging an agent conversation from the command line.
In 1 T3 Code release, first seen Oct 22Add a line to your agent's skill (a reusable instruction file) telling it to stop a slow helper model and get back to work when it is stuck waiting and has other work to fix. This keeps your agents from sitting idle.
In 1 video, first seen Nov 13Try a proposed code change (PR) yourself before merging, from any of your devices. A thread checks out the PR and starts a dev server shared over Tailscale, a private network between your devices. For a desktop app, a PR label can produce a downloadable preview build.
In 1 video, first seen Sep 8Ask for everything in one prompt: build the feature, file a PR (a proposed code change), spin up a test copy you can try, and babysit the PR, meaning keep watching it and fixing issues. The next time you look, you can test and merge.
In 1 video, first seen Oct 1Add one line to prompts that gives Opus permission to push back: 'If you think this is a good idea, do it. Otherwise, suggest something else.' Opus rarely challenges the task it's given otherwise.
In 1 video, first seen Nov 12Judge a security bug by two things: how long until everyone actually runs the fix, and how much access the attack gives. A bug in your own server is fixed once; one in software users install or in a library you rely on stays open until everyone updates.
In 1 video, first seen Oct 22After a first round of fixes, ask the agent to dig deep and come back with several extreme proposals, explicitly allowing it to "boil the ocean". Use it when normal fixes have stalled.
In 1 video, first seen Oct 5A prompt with layers: state the goal as non-negotiable, offer your own idea 'in my opinion' so the model can push back, admit what you don't understand, and ask for an audit rather than code, with permission to stop early if it finds a simple fix.
In 1 video, first seen Sep 8Swing between building features and improving the foundation. When a feature is harder than it should be, stop, make the codebase easier to build on, then return to the feature.
In 1 video, first seen Nov 9A hands-on way to do performance work: name one slow thing, have the agent list theories, build demos of each, then try them yourself to see if it actually feels faster.
In 1 video, first seen Oct 5Every long task prompt states three things: what you want, what counts as done, and when the agent should stop and ask you. The last part lets the agent stop when it is genuinely stuck.
In 1 video, first seen Sep 24A way of setting up your codebase so a very capable developer who has never seen it can do useful work right away. Each agent session starts with no memory of earlier ones, so every thread is effectively a new hire.
In 1 video, first seen Sep 6Cut slow agent actions besides writing code, such as git commands, builds or browser snapshots, when the model itself is fast. With a fast model, those steps can double your total wait.
In 1 video, first seen Oct 6Collect every time your coding agents failed, such as a broken tool call or a wrong path through your code. Turn each one into an eval, a repeatable test of whether the agent gets it right, so you can measure how well your codebase steers agents.
In 1 video, first seen Nov 5Before a big change, ask agents to review your architecture or planned changes and to investigate theories you want to test. You keep the strategic decisions and use the agents as advisors. Once you have a confident plan and a way to check the result, they carry it out quickly.
In 1 video, first seen Aug 24When you take over code nobody remembers, ask an AI questions about it and have it check its answers against how the code actually behaves. This helps you get up to speed before changing anything.
In 1 video, first seen Sep 6Set up extra agents on crons (timed, recurring jobs) that check whether a long-running agent has stalled waiting on a question. They notice when the main agent is stuck so it can be unblocked.
In 1 video, first seen Nov 18CLAUDE.md is the instruction file Claude Code reads in every session. Anthropic suggests a rule for it: keep going when no input is needed, and stop only when blocked or before anything destructive. This cuts needless pauses on long runs.
In 1 video, first seen Sep 24Wispr Flow turns speech into text on screen. Theo pairs it with a small desk podium mic, about $70, that picks up even whispers, so he can dictate quietly around other people.
In 1 video, first seen Sep 15Format a Linux drive as XFS with VDO underneath, a layer that compresses data before it hits the disk, using the lightweight LZ4 compressor. Duplicate files across agent work trees then take little real space, and work trees and installs get fast.
In 1 video, first seen Aug 20Theo Daily is unofficial and not affiliated with Theo or t3.gg. Written by Claude Opus 5.5 from each video's transcript.