Google Team Starts Writing Documentation for AI: Three .md Files Keep the Coding Agent in Check
In a Google Cloud Tech video, Annie Wang and Tilde Thurium break down three Markdown specification files: DESIGN.md, AGENTS.md, and SKILL.md. They're not documentation for humans—they're for agents.
Google Cloud Tech released a video in which Annie Wang and Tilde Thurium talk about how the team is rolling out agentic coding practices. The core isn't some new model; it's three Markdown files:
DESIGN.mdAGENTS.mdSKILL.md
According to the video, these spec files define how your agent should think, how it should design, and how it should execute tasks in your codebase.
Original post link: goo.gle/4f0Q5H1
This reply puts AGENTS.md very plainly:
wait so AGENTS.md is basically a readme for the robot, love that
A README for the robot. The analogy holds. When humans enter a new project, they first check the README to figure out how to run things, where the code is, and what pitfalls to watch out for; when an agent enters a codebase, what it reads is AGENTS.md.
What's more interesting is DESIGN.md. Fernando Ruiz's reply hits the pain point:
DESIGN.md is the one I'm most curious about. Spacing rules and component decisions usually live in a Figma file the agent never sees, so writing them down for it makes a lot of sense.
Things like spacing rules and component choices usually sit in Figma, where the agent can't see them at all. If you ask it to change a button, it doesn't know that this project's 8px grid is an ironclad rule, nor which card to use in which scenario. Writing these into DESIGN.md is equivalent to laying out the tacit constraints in a designer's head as text the agent can read.
The division of labor among the three files is roughly this: DESIGN.md handles visual and component decisions, AGENTS.md handles project conventions and collaboration methods, and SKILL.md handles execution steps for specific tasks. They're all Markdown, plain text, checked into the repository, and travel alongside the code.
What's worth noting here is the direction. Over the past few years, everyone has been busy feeding models more context and longer windows; now some people are starting to seriously write “documentation for agents.” The audience for documentation has changed, so the writing has to change too: docs for humans can omit premises, while docs for agents must spell the premises out.
As for whether this can really scale to an entire team, the video doesn't give a conclusion. But at least it acknowledges a reality: agents won't automatically know what your project looks like—you have to tell them.
发布时间: 2026-11-19 01:29