Claude Code 3.0 (New Mods Feature): Anthropic has SILENTLY dropped their BIGGEST NEW Feature!

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Claude Code 3.0 (New Mods Feature): Anthropic has SILENTLY dropped their BIGGEST NEW Feature! · Channel: AICodeKing · 07:07 · Watch on YouTube

Topics: AI agents, Task and workflow automation, AI evaluation and benchmarks, Claude Code, AI developer tools

Summary

Claude Code now supports AGENTS.md instructions, while Anthropic presents mods as a way to customize how Claude Code works. The video explains how shared project instructions interact with Claude-specific files and covers built-in mods, including a live diff panel. It also reviews redesigned cloud projects with coordinated threads, background computer use, session continuation, and Fable 5.1 support. Finally, it describes tools for checking skill context costs and evaluating whether plugins improve results, while noting that some features are limited or still in development.

Key points

Version 2.1.277 added support for AGENTS.md, used by default when a project has no Claude-specific instruction file; this setting can be changed in /config.
A project’s Claude-specific instruction file takes priority by default, though users can choose to load both files.
Mods use function hooks to customize Claude Code behavior; custom mods are early access, and their API may change.
The live diff panel shows changed files and line counts, and lets users select lines to include in a message; it requires a Git repository and a terminal at least 110 columns wide.
The redesigned projects beta uses a coordinator to delegate work to separate cloud-session threads that share project memory and have their own branch and repository copy.
Background computer use lets Claude work in approved Mac apps while the user continues using the computer; the desktop app must stay running.
The /res command can continue a terminal session in the desktop app with its conversation and context intact.
Fable 5.1 is supported in Claude Code from version 2.1.257 and has a 1 million token context window; gateways may require its explicit model name.
/skill Doctor reports skills’ context cost and usage, while /plugin eval tests a plugin against test cases and can compare results without it; evaluations consume usage.

Chapters

Start	Chapter	Summary
00:02	AGENTS.md support and project instructions	Claude Code can read AGENTS.md, and the video explains how it works alongside Claude-specific project instructions.
01:39	Mods and the live diff panel	The video introduces mods as a way to customize Claude Code and describes the built-in mods and live diff panel.
03:17	Coordinated projects and background computer use	It covers cloud projects that delegate work across threads, as well as desktop background computer use.
04:24	Desktop sessions and Fable 5.1	The video explains continuing terminal sessions in the desktop app and outlines Fable 5.1 support and access considerations.
05:32	Skill and plugin evaluation	It describes checking skill context costs and testing whether plugins improve outcomes before adopting more complexity.

Quotes

Claude Code can now read agents.mmd. — 00:02

Mods let developers hook into that behavior with code. — 01:39

A project now has a coordinator that can break work into separate threads, delegate tasks, and review the results. — 03:17

The results help you see what the customization contributes. — 05:32

Concepts and entities

Claude Code, Anthropic, AGENTS.md, CLAUDE.md, mods, Projects, Fable 5.1, /skill doctor, plugin eval

Full transcript (original language)

00:02 [music] Hi, welcome to another video. Claude Code can now read agents.mmd. So if your project already uses that file to tell coding agents how to work, Claude can finally pick up those instructions directly. But the part that caught my attention is how Anthropic built this. Agents.md support is a built-in mod, and mods are their upcoming way to let you customize Claude Code itself. There are some other substantial updates too, including projects that coordinate work across multiple sessions, background computer

00:32 use, and tools for checking whether your plugins help. I’m covering August 20th through September 18th, 2026. First, the agents.m MD update arrived on September 18th in version 2.1.277. By default, Claude Code uses it when the project has no claude specific instruction file of its own, such as cloud MD. You can change this under project instructions in /config. At launch, Anthropic says, “This support isn’t available on Bedrock, Vert.Ex, or Foundry yet.” An [clears throat] instruction file is where you might explain how to run tests, which package manager to use, or which folders contain

01:07 generated files. If you’re switching coding agents on the same project, sharing those rules can save you from maintaining duplicate copies. If your project already has its own clawed MD, those instructions take priority under the default setting. There is also an option to load both files. That gives you a way to keep shared project guidance alongside Claude specific instructions. I’d keep the shared file focused. Put the commands and conventions the agent actually needs in there, then check its work. Being able to read a file doesn’t guarantee that every instruction will be followed

01:39 perfectly. Now, the mods part is more interesting to me. The harness is the software around the model. It handles things like assembling context, calling tools, and displaying the session. Mods let developers hook into that behavior with code. Anthropic defines a mod as a plug-in that uses function hooks and its early demos include changing interface components and filtering tool output before the model receives it. That explains why agents support is a useful first example. The customization changes which project instructions reach the

02:10 model. The also says developers will be able to build their own versions of that instruction loading behavior. The public repository now contains four built-in mods. the agents domain loader, the diff panel, telemetry, and an organization security mod. You can inspect how those features are implemented. Custom mods are still early access. The documentation says function hooks must be enabled, and the API can change between releases, so the built-in agents MD feature has shipped while the broader customization system is still developing. For someone building a specialized coding workflow, that’s

02:43 worth watching. I’d be interested in a mod that assembles the right instructions for each part of a large project without loading every team’s rules into every task. The diff mod also connects to a practical update from September 3 in version 2.1.260/de gained a live panel beside the conversation in full screen mode. You can watch the changed files and line counts update as Claude edits or runs commands and select lines to attach to your next message. It needs a git repository and a terminal at least 110 columns wide. That makes feedback easier. If Claude changes the wrong

03:17 function, you can point at the actual edit and explain what needs to change while the conversation stays visible. The other big announcement is the redesigned projects beta from September 17th. A project now has a coordinator that can break work into separate threads, delegate tasks, and review the results. Those threads share project memory and each runs as a clawed code cloud session with its own branch and repository copy. Because the work runs in the cloud, it can continue after you close your laptop. For example, I would try giving it a feature that needs an API change and a front-end update. The

03:50 useful test would be whether it coordinates those pieces and catches mismatches between them. Access is limited initially to selected Pro and Mac subscribers using cloud sessions who don’t already have projects on web or desktop. Team and enterprise access and local execution are coming later. Parallel threads can also use your allowance faster. I like the direction, but I judge it on how well it handles the handoffs. Getting several agents to start working is only part of the job. Desktop users also got background computer use during the week of August 31 through September

04:24 4th. On Mac OS 15 or later, Claude can work in approved apps while you keep using your computer. It’s a beta for Pro and Mac subscribers. That could help with tasks in a native app that has no useful command line interface. You approve the apps Claude can access and the desktop app needs to stay running. This is work happening on your computer, so keep it awake while Claude work. Another useful desktop improvement arrived in the August 24 update week. You can use slashres in the app to continue a session started in the terminal with its conversation and context intact. On the model side, Fable

04:59 5.1 launched on September 4th. Claude code supports it from version 2.1.257 with a 1 million token context window/model Fable selects it in the normal setup. Claude apps gateways have an exception. Their alias still points to Fable 5. So a gateway serving 5.1 needs the explicit model name. Check your access and usage requirements before switching. Depending on your plan, Fable can use separately build usage credits. Finally, two additions fit nicely with all this customization. /skll Doctor shows the context cost and

05:32 usage of your skills, helping you identify ones you might want to disable. That matters because the listing of available skills occupies context even when those skills aren’t invoked. This arrived in the early September update window. Then version 2.1.269 released September 11th added clawed plugin eval. You give it test cases and scoring checks and it runs them with your plugin. By default, it also runs a comparison without the plug-in. The results help you see what the customization contributes. These are

06:02 real model calls, so evaluations consume usage. That’s something I’d use before adding another complicated workflow to my setup. Pick a few tasks you repeat, define what a good result looks like, and see whether the plug-in improves the outcome enough to justify the extra cost and complexity. The update I’m most interested in is still mods. Agents on MD solves an immediate compatibility problem, and it gives us a concrete example of what changing the harness can do. I’d like to see what people build once that interface settles down. Let me

06:32 know which update you’d actually use. Would you customize Claude code with mods or are you more interested in having projects coordinate the work for you? Overall, it’s pretty cool. Anyway, let me know your thoughts in the comments. If you like this video, consider donating through the [music] super thanks option or becoming a member by clicking the join button. Also, give this video [music] a thumbs up and subscribe to my channel. I’ll see you in the next one. Until then, bye.

07:07 >> [bell]