Claude Code Opens the Door to Shared Briefs

The awkward part of adding a second coding agent isn’t choosing the tool. It’s briefing it. One agent reads the repository’s instructions, another starts with a blank page, and the team quietly maintains several versions of the same “here’s how we work” prompt.

That changed this week. In Claude Code 2.1.277, released September 18, Claude Code will read `AGENTS.md` when a repository does not contain `CLAUDE.md`. The fallback gives a project one place to describe its working rules while leaving room for different coding tools to use the same brief. Anthropic’s changelog, reported in the release note, also says the change speeds startup for scripted sessions. The product consequence is straightforward: instruction management can move out of each agent’s custom prompt and into a shared project surface.

The brief becomes infrastructure

Picture a team with a repository and two agents. The first knows which tests to run, which directories are off limits, and how the team names a pull request. The second knows only what its operator remembered to paste into the session. That is not a model-quality problem. It is a documentation and governance problem wearing a model-quality costume.

`AGENTS.md` gives product teams a cleaner default for that handoff. Claude Code now checks for its own `CLAUDE.md` file first, then uses `AGENTS.md` when that file is absent. The distinction matters because the fallback is predictable. A team does not need to create a Claude-specific instruction layer before another agent can understand the repository’s expectations. The Claude Code changelog records the change as a Sept. 18 update, and the discussion linked there describes the behavior plainly: Claude Code reads `AGENTS.md` if there is no `CLAUDE.md`.

That makes the file more than a convenience. It becomes a contract between the repository and the agents operating inside it. The contract can cover the things a product team repeatedly explains in tickets, onboarding notes, and chat: what a safe change looks like, which checks matter, and where an agent should look before it edits. The evidence here doesn’t tell us how teams will govern those rules, or whether every agent will interpret them identically. It does give teams a stable place to start.

Stop prompting from scratch

The second-order effect is a change in what “good prompting” means for coding workflows. The old instinct is to improve the session prompt: add another instruction, paste in another example, remind the agent of the repository’s quirks. That can work for one session. It decays when the next operator, script, or coding agent enters the same project.

A shared file changes the maintenance unit. Instead of asking every operator to remember the right setup, you can review the project brief as part of the product’s operating rules. Monday’s practical move is small: open the repository, create or review `AGENTS.md`, and write only the instructions an agent needs to act safely and consistently. Then ask one person to own changes to that file, just as they would own a product requirement or release checklist. Test Claude Code against it, especially in a scripted session, and watch for instructions that are vague enough to invite different readings. Keep the file about durable project behavior, not a transcript of one agent’s preferences.

Anthropic also says this release retires the `TaskOutput` tool, but the useful product signal is the fallback itself. Claude Code is making the project’s existing documentation structure more portable, which reduces the pressure to build a bespoke prompt system for every tool. The work doesn’t disappear. It moves to a more reviewable place: deciding what every agent should know before it touches the code. That is the governance task worth putting on the roadmap now, before every agent gets its own version of the brief.