Why one AGENTS.md beats forty.

The cost isn't visible until you go looking

None of this shows up as an incident. It shows up as friction that everyone in the org has learned to route around: a reviewer who knows repo A uses a different test command than repo B and mentally translates every time they context-switch, an agent that confidently runs the wrong build command because the file it read was written before the build system changed, a new hire who reads one repo's AI instructions and assumes they generalize. Every one of these is small. All of them together is why AI-assisted review takes longer than it should, and why nobody can say with confidence what "the standard" actually is.

One canonical file, compiled per tool

baselane's answer is to stop treating each repo's agent instructions as something to author by hand at all. A repo declares its commands, constraints and conventions once, in one canonical spec. From that single source, baselane compiles the format each tool actually expects: an AGENTS.md for the emerging cross-tool convention, a CLAUDE.md that includes it, Cursor rules, Copilot instructions — same content, different shape, because every tool insists on its own file and its own front matter.

Here's roughly what that compilation looks like for the two Claude-facing files in this repo, for instance:

AGENTS.md        # canonical: commands, conventions, where things live
CLAUDE.md        # @AGENTS.md — a one-line include, not a fork
.cursor/rules/*  # compiled from the same source, Cursor's format

Nobody edits CLAUDE.md directly and lets it drift from AGENTS.md. There's one place a convention gets written down, and every tool-specific file is a compiled view of it.

Content-preserving, always

The one hard constraint on this compilation step: it must never clobber what a developer already wrote by hand. If a repo has custom notes in its CLAUDE.md outside baselane's managed region, those survive every update untouched. The compiler merges its managed section in and leaves the rest exactly as it found it. A standard that silently overwrites a team's own documentation isn't a standard anyone will trust twice.

What one file actually buys you

• A reviewer who's onboarded to one repo's conventions is onboarded to all of them — the commands, the file layout expectations, the constraints are declared the same way everywhere.
• An agent working across repos gets the same shape of context every time, instead of re-learning conventions per repo from whatever ad hoc instructions it finds.
• A platform team gets one place to improve the standard — a change to the canonical spec propagates as a reviewable PR to every subscribed repo, instead of forty manual edits.

None of this requires giving up per-repo differences that are real — a repo genuinely on pytest should say pytest, not pretend to be a Vitest repo because that's what the template said. The point isn't uniformity for its own sake. It's that the shape of the declaration is identical everywhere, so the differences that remain are the ones that actually matter.