The trim numbers are worth doing by hand. Anthropic's doctor command put posthog.com's AGENTS.md at roughly 1,780 resident tokens and recommended cutting about 350 of them, a fifth of that one file [7][17]. Set that against the roughly 6K tokens per session the same report expected to save once three unused plugins and three skills were switched off [8]: the memory-file prose accounts for under 6% of the saving [18]. The document that gets argued over in code review is the small line item here.
Teams treat AGENTS.md as a briefing for the agent, but in practice it becomes a changelog of last month's workarounds. PostHog's merge queue line is the clean example: correct when it was written and false within days, in a way no tool in the repo can see, because the queue's paused state lives in a GitHub setting rather than in code [10][12]. Anthropic's version of doctor works from what it can derive from your code, and correctness is not on that list [9].
Subtraction needs a regression net, or it turns into adding the line back after every incident. PostHog's net runs its onboarding wizard against about 40 sample apps in CI, opens one PR per app, and has a second agent called pr-evaluator grade the diffs and session logs rather than merging them [15]. Gotchas seen in real wizard runs get written down as framework-specific instructions and folded into those evals [14]. The real price of a lean context file is that the knowledge has to live somewhere that fails loudly. The same team removed a default that pointed its scripts at the repo root, because models now infer monorepo structure well enough that headless runs can follow the model's own recommendation [16].
For Monday, two questions per line, and four boxes. Can the model work this out from the repo? And does the underlying fact live inside the repo, where a tool can check it? Model-inferable and repo-checkable is a delete, and doctor will usually flag it for you. Not inferable but repo-checkable earns its place, and CI can verify it. Not inferable and not repo-checkable is the quadrant PostHog's merge queue line sat in: keep it if you must, but attach an owner and an expiry, because nothing in your toolchain will notice when it goes false. Inferable but not checkable goes first, being both redundant and prone to rot.
The tradeoff sits in the same paragraph as the recommendation: prune hard and you will rediscover a gotcha the expensive way, in a review at 6pm on a Friday. Rewriting the file by hand, as Theo reported doing, is the maximal version [5]. The milder one PostHog suggests is to run doctor after each upgrade, then read the file yourself, and delete any line whose prevented failure you cannot name out loud [13]. That test is doing more work than the token count, because a line you can justify is also a line you will remember to update.
