Everything I promised in the reels.

• REA on GitHub (morluto/rea)

  The official repo. Its README warns that copies exist: get REA only from this repo, rea.tools or the rea-agents npm package.

• rea.tools

  REA's own site, with the Calculator and dinosaur game examples from the video.

• AI LABS: 5 ways to use REA

  The billing page test from the video is in this one.

npx rea-agents setup

Pick Claude Code (or Codex, Cursor and others), approve the changes, then restart your agent. Then ask it something like: Understand how search works in the Notes app, show me the evidence, and build a similar feature for my project.

Use it on software you own or are allowed to examine. The analysis runs on your own machine. I have not run REA myself.

Since Claude Code v2.1.299, the /usage command warns you about the four things that eat your limit fastest. Run /usage first, then fix the warning at the top.

• Always-on context: your CLAUDE.md loads every session. Move rarely used instructions into skills, which load only when used. Keep CLAUDE.md under 200 lines.
• Max effort: thinking can run to tens of thousands of tokens per request. For simple tasks, lower it with /effort.
• Large tool results: a 10,000-line log stays in your context. A hook that returns only the error lines cuts it from tens of thousands of tokens to hundreds. A subagent can also read it and bring back a summary.
• Workflows: one run starts many agents. Past 25 agents or 1.5 million projected tokens, Claude Code shows a Large workflow warning. Try a workflow on one folder first and watch /workflows.

/usage
/effort
/context
/mcp

/usage shows the warnings. /effort sets how hard Claude thinks. /context shows what takes space in the conversation. /mcp lets you turn off MCP servers you don't use. Press d or w in /usage to switch between the last 24 hours and the last 7 days.

Bonus: a 200K+ token conversation left idle for about 54 minutes is now compacted before its one-hour prompt cache expires, so your next message doesn't reprocess all of it. You can turn this off with the Idle compaction row in /config.

• Manage costs (Claude Code docs)

  The /usage breakdown and every fix above.

• Workflows: Cost (Claude Code docs)

  The Large workflow warning: 25 agents or 1.5 million tokens.

• Claude Code changelog

  v2.1.299, 14 October 2026: the /usage warnings and idle compaction.

All of this is from Anthropic's changelog and docs. I have not run these on my own limits.

Two commands for moving to Claude Code from Codex, Cursor or Gemini CLI. The first brings your setup over. The second makes the setup fit how you work.

/import cursor --dry-run
/import cursor

Use codex, cursor or gemini. It brings instruction files, MCP servers, commands, subagents and skills. Your instructions are copied into your CLAUDE.md. --dry-run shows the list without writing anything. Anything it cannot map, it lists, and it can offer to write a skill that finishes those with you (that is what AI LABS saw on their Codex setup).

CLAUDE_CODE_NEW_INIT=1 claude
/init

With the variable set, /init asks which things to set up (CLAUDE.md, skills, hooks), reads your code, asks follow-up questions and shows you a plan before it writes any file. You can also put the variable in the env block of your Claude Code settings so it stays on. /import needs Claude Code 2.1.213 or later (2.1.265 for Cursor) and is not available on Amazon Bedrock, Google Cloud's Agent Platform, Microsoft Foundry or Claude Platform on AWS. I have not run these on my own setup.

• Claude Code docs: Commands

  The /import and /init rows.

• Claude Code docs: Migrate instructions from other tools

  What /init and /import read and where it ends up.

• AI LABS: Claude Code Features That 90% Don't Know Exist

  Both commands running, from 1:04.

• Keep Claude working toward a goal (/goal)

  Claude Code's own docs for the loop in the video.

• karpathy/autoresearch

  Karpathy's loop: the agent edits one file, a locked file scores it, keep or undo, repeat. Read program.md.

• Karpathy's results post, 9 March 2026

  About 700 tries in two days, about 20 changes kept, and an 11% faster time to GPT-2.

• AI LABS: their Claude Code loop, step by step

  The restaurant example with 11 checks, and the locked checks folder.

Step 1. Ask Claude to write the checks first, before any code: tests that prove the feature works for a real user, not only the rules behind it. Read them and add what is missing.

Step 2. Lock the checks so the agent cannot edit them. Put them in one folder and add a deny rule to .claude/settings.json in your project. This is the rule AI LABS showed in their video:

{
  "permissions": {
    "deny": ["Edit(checks/**)"]
  }
}

Step 3. Start the loop with /goal. Claude works, a second model checks the condition after every turn, and Claude keeps going until it is met.

/goal all tests pass and no test file is changed

To stop it early, type /goal clear. To cap it, add a limit to the condition, like "or stop after 20 turns". The checker only reads what Claude shows in the conversation, so ask for a check it can prove, like the test command's output. I have not run this setup myself.

• Redesigning source control at Anthropic

  Anthropic's own post, 13 October 2026. All the numbers in the video are from here.

• git sparse-checkout

  Git's own docs for the trick.

Scion is internal to Anthropic, so you can't install it. The idea behind it is already in Git: only fetch the folders you need. These are the two commands.

git clone --filter=blob:none --sparse <your repo url>
cd <repo>
git sparse-checkout set <folder you need>

The first command clones without downloading file contents up front and starts with only the top folder. The second tells Git which folders to fill in. Anthropic says they used Git's own tools like this before Scion, and that it bought them time. I have not run these on a repo of their size.

• Anthropic's launch video

  One comment on a chart turns into a 38 second film.

• Get started with Claude Motion

  Anthropic's help page.

• Higgsfield Katana

  A video editing tool inside Claude, built on Claude Motion.

• claude-motion-design, a free Claude Code skill

  Open source, for people without a Team plan. It makes this kind of video the same way, as code. I have not run it myself.

For now Claude Motion is in beta on Team and Enterprise plans.

Start from your own problem:

1. You keep pasting the same instructions into chat: that is a skill.
2. You keep copying data from a tab Claude can't see: that is an MCP server.
3. Something must happen every time: that is a hook.
4. You want a new panel or a button on your screen: that is a mod.

• The comparison table

  From the Claude Code docs.

• When to add what

  The longer list, also from the docs.

• claude-code-setup, Anthropic's official plugin

  It reads your project and recommends which skills, hooks and MCP servers to add. It does not cover mods.

/plugin install claude-code-setup@claude-plugins-official

recommend automations for this project

Boris Cherny created Claude Code. He does not use a prompt template. He tells Claude three things:

1. What you want it to do
2. How much effort you want it to spend
3. How it should verify that it did the right thing

A line from his own prompt, which is how he sets the effort:

Make it really good, use lots of tokens and iterate till you're proud of it.

• His post on X

  @bcherny, 6 October.

• Token Weather, from Anthropic

  A line above your prompt that shows how full your context window is.

• image-view, by Jarrod Watts

  Paste a screenshot and see a thumbnail of it. Needs a terminal that can draw images, like Ghostty or kitty.

• mission-control, by Hamza Zafar

  A live pane with every subagent, every tool call and every file your agent touches.

/plugin marketplace add jarrodwatts/claude-image-view
/plugin install image-view
/reload-plugins

Each one installs with two or three commands. The commands for the other two are in their README. You need Claude Code 2.1.287 or later.

1. Update to Claude Code v2.1.252 or later.
2. Type /skill-doctor and read the last column.
3. Open /skills and turn off the ones marked never.

The report lists every skill with what it costs, how often you use it and when you last used it. It opens in the Stats tab of the plugin manager.

• Skills in the Claude Code docs

  The official page.

• Anthropic's three example mods

  Blast Radius, Token Weather and Replay Theater. Look under claude-code/mods.

• Mods in the Claude Code docs

  What a mod can do and how it runs.

/plugin install <mod-name>@<marketplace>

claude plugin validate ./<mod-folder>

A mod is code running with your permissions, and it is not sandboxed. Only install mods from people you trust. You need Claude Code 2.1.287 or later.

npx claude-mem install

• The installer asks you to sign in to their hosted service. You can skip it and use your own Claude plan or API key.
• Anything you wrap in private tags is never stored.
• It works with Claude Code, Codex, Gemini, Copilot and OpenCode.

• claude-mem on GitHub

  Free and open source.