Where Every AI Coding Tool Looks for Your Instructions

Eighteen AI coding tool families read at least twenty-one different standing-instruction file paths, and only seven of them publish an explicit precedence order for what happens when several apply at once. That gap is the reason this page exists. Below is the complete map: the file or files each tool reads, the scopes it searches, whether it publishes a conflict rule, whether the content loads on every request or only when triggered, and the vendor documentation page behind every cell.

One boundary before the data, because a companion piece published the same day covers the neighboring question. This post is a map of where the files live and who reads them. It makes no recommendation about what belongs in them and publishes no measurement of our own. What to cut, and what an audit of a real instruction estate actually finds, is the subject of our record of the files we cut and what broke. The only numbers here are limits other people published. Every file path, scope, precedence rule and byte cap comes from a vendor documentation page, retrieved 2026-08-30, and where a vendor is silent the cell says so and stays visible.

Two related layers are also out of scope. What makes an instruction file bad — the named configuration smells and how often they occur across open-source repositories — belongs to our config-smells reference. And prompt technique is a different layer entirely: standing instruction files are configuration, loaded before anyone types; prompts are per-request.

The reader's problem does not need a hypothetical. The repository this post was written in carries six committed instruction surfaces at once, all six tracked in git, measured with wc and git ls-files on 2026-08-30: a root CLAUDE.md at 101 lines, a root AGENTS.md at 42 lines, eight scoped per-directory CLAUDE.md files totalling 58,049 bytes, seven .cursor/rules/*.mdc files totalling 54,105 bytes, a .claude/skills/ directory, and a .github/copilot-instructions.md at 127 lines.

Nobody designed that. It accreted one tool at a time, and each file was a reasonable thing to add on the day it was added. The problem is not the count. It is that four different tools, pointed at this one repository, read four different subsets of those six surfaces — and nothing in the repository tells you which subset belongs to which tool. The table below is that mapping for our own files, and it is a worked instance of the general map in section 03.

Read across the AGENTS.md row. One tool ignores it, one tool reads nothing else, one reads it as one rule type among four, and one reads it as its lowest-priority repository tier. Read down the Claude Code column and three of the six surfaces are invisible to it at runtime. There is no arrangement of six files that makes that table uniform, which is why the answer to “where do I put my instructions” is a lookup and not a convention.

And then there is the row we did not expect to write. Building this table surfaced a real misconfiguration in our own repository, and it is the strongest argument available for the table existing, because the mistake is invisible without one.

Somebody wrote path-scoping into a file with no documented path-scoping mechanism, it sat there through months of ordinary work, and it was found by reading the vendor's reference table column-by-column against our own tree. That is what a map is for. The fix is booked; the row is published as it stood on the retrieval date.

The headline story of the last two years has been convergence on AGENTS.md — a single cross-tool filename, now stewarded by the Agentic AI Foundation under the Linux Foundation. The marketing answer to “who supports it” is “almost everyone.” The documentation says something more interesting, and it is the single most consequential fact in this map.

That is a vendor documenting a deliberate non-behavior, in one sentence, with two documented bridges attached: put @AGENTS.md at the top of your CLAUDE.md, or symlink one to the other with ln -s AGENTS.md CLAUDE.md — and on Windows the symlink needs Administrator or Developer Mode, so the import is the portable option. It matters more than any other cell in the table because Claude Code is the tool most likely to be running in a repository that already has an AGENTS.md.

Now look the other direction. CLAUDE.md is read by more non-Anthropic tools than AGENTS.md is read by Anthropic's. Cursor reads it — and treats it as “always applied to every conversation, regardless of any alwaysApply frontmatter setting.” Zed has it eighth in its ordered filename list. Amp uses it as a per-directory fallback when no AGENTS.md is present. Devin CLI treats it identically to AGENTS.md. Devin Desktop's importer pulls it in. Factory Droid lists it in its compatibility set. Copilot's cloud agent and CLI read it as agent instructions. Seven tools, one filename, and no arrow back.

The practical consequence for anyone running more than one tool: a repository with only AGENTS.md is silently unconfigured for Claude Code, and a repository with only CLAUDE.md is read by most of the field but ignored by Codex, whose documentation states that filenames not on its list “are ignored for instruction discovery.” Neither failure produces an error message. Both produce an agent that behaves as though your conventions do not exist.

Eighteen tool families, nine columns, one row each — plus a superseded Cursor file kept visible because it still wins in Zed, and a final row for the AGENTS.md convention itself, which is a specification rather than a tool. Every path, scope, precedence rule, load behavior and size limit in this table came from the vendor documentation page in the last column, retrieved 2026-08-30. Cells that read not documented mean we read the vendor's rules or memory page and it does not address the question. They are not gaps in our reading; they are the published state of the ecosystem.

The table scrolls horizontally. If you cite one cell, cite its retrieval date with it — three of these vendors changed a brand or a default filename in the weeks before we read them.

Two footnotes on rows that will surprise people. Warp requires the filename in all caps — agents.md and Agents.md are not recognized — while Factory explicitly reads both of those variants. Two tools, opposite behavior, one file. And Codex's row covers discovery, ordering and its byte cap only; what happens to those instructions after they are loaded, including how they are stored and whether the loaded set is auditable, is the subject of our piece on Codex's encrypted agent instructions.

A yes-or-no column for AGENTS.md support would be a lie for at least five of these tools. Copilot's answer depends on which surface you are in. Gemini CLI and Aider read it only if you add a line of configuration. Warp reads it by default but lets WARP.md beat it in the same directory. Zed calls it the primary file and then ranks it seventh out of nine. So the honest presentation is four tiers.

Continue is the tool deliberately absent from all four tiers. Continue's CLI documents /init as creating an AGENTS.md for the current project, which proves it writes one — and no page on its rules deep-dive or its config.yaml reference says it reads one as standing instructions. Putting Continue in the “yes” column on the strength of a scaffolding command would be exactly the kind of inference this table exists to avoid, so its cell reads not documented.

This is the column readers arrive looking for and the column most vendors do not fill. Seven of eighteen publish an explicit ordering for what happens when several instruction files apply at once. Three more publish a scope-pair rule — workspace beats global — and stop. Five publish a load order, a merge mechanic or a “mental model” without naming a conflict winner. Three publish nothing, or have no chain to rank.

That distribution is itself the finding. Most tools tell you where the file goes. Fewer tell you which file wins.

Among the seven that do document an ordering, the mechanisms are not variations on one theme. They are three genuinely different algorithms, and knowing which one your tool uses is worth more than memorizing its file paths — because it tells you what to predict.

One consequence worth stating plainly, because it is the most common way a reader gets burned: in a concatenating tool, a file you forgot about does not lose. It gets appended. In a first-match tool, a file you forgot about does not get appended — it wins outright and silences everything below it. Neither failure emits a warning, and the only reliable defense is knowing which list your tool reads and in what order.

Three tools publish a way to check. Claude Code's /context lists loaded memory files and its InstructionsLoaded hook logs exactly which instruction files loaded, when and why. Gemini CLI's /memory show prints the full concatenated hierarchy. Amp exposes agents-md list in its command palette. Warp marks which rules fired under References in the conversation. If your tool publishes one of these, use it before you trust any table, including this one.

Ten of the eighteen tool families publish no size limit of any kind. Four publish a number, and no two of them agree on what to measure. The same job — bounding how much standing instruction text enters a model's context — is expressed in kibibytes, characters, lines and mebibytes depending on which vendor you ask.

The incoherence is the point, and it has a practical edge. A repository sized to Devin Desktop's 12,000-character per-file ceiling is comfortably inside Factory's 80,000 and nowhere near Claude Code's 4 MiB — but Codex's 32 KiB applies to the combined chain, so a monorepo with a root file and six nested ones can breach Codex while every individual file looks small. If you maintain instructions for several tools, the binding constraint is whichever vendor measures the thing you are not measuring.

Note also that four vendors who publish a cap separately advise staying far below it. Factory: “These are caps, not targets. Smaller files are usually better.” Claude Code pairs its 4 MiB skip threshold with a 200-line target. Devin CLI is the bluntest: “Rules and AGENTS should be kept as small as possible.” A published limit is not a budget.

This is the column that costs money, and it is the one most summaries of this topic omit. An always-loaded file enters the model's context on every request in scope, whether or not it is relevant. An on-demand file does not. Same content, same repository, different bill — and the academic evidence that convention files raise inference cost materially is covered in our read of the ETH Zurich study on AGENTS.md and inference costs. The mechanics of budgeting for it live in our agent token-budget framework.

The column resists a binary answer because most tools are both, and which one a given file gets is often decided by the file's own frontmatter rather than by the tool. Four patterns cover the map.

Two rows carry a trade-off sharp enough to change a decision. Kiro's: adopting AGENTS.md removes your ability to make the content conditional, because the file has no frontmatter — “AGENTS.md files do not support inclusion modes and are always included.” Portability costs tokens, and one vendor says so outright. And Claude Code's: splitting a long file into @path imports “helps organization but doesn't reduce context, since imported files load at launch.” The reorganization that feels like a saving is not one.

Which is where this post stops and its companion starts. Everything above is a map of where content lives and when it loads. What to remove from it, what a two-model audit of a real instruction estate actually finds, and why a headline percentage from a vendor does not transfer to your repository are the subject of our own instruction-file audit and what it broke. That post publishes a measurement with a denominator; this one publishes vendor-documented caps and nothing of our own. Same day, same layer, deliberately different jobs — use this page to find the files, and that one to decide what stays in them.

These are kept visible on purpose. A reference that shows its gaps is more citable than one that fills them, and every row below is a question we went looking for an answer to and did not find on a vendor page. We did not infer any of them from observed behavior, changelogs or source code.

One class of gap is more interesting than the rest. Where a vendor publishes a load order but no conflict winner — Gemini CLI, Amp, Devin CLI, Continue, Factory — the documentation is describing a mechanism without committing to a guarantee. That is not sloppiness; concatenation makes “which one wins” a genuinely fuzzy question, because both files are in the prompt and the model decides. But it means a reader cannot predict behavior from the docs, and it is why those cells say not documented rather than “the later one wins,” which is a guess dressed as a rule.

The whole value of a reference table is that a stranger can check it. Here is exactly what was collected, from where, on what date, what was left out, and what the table cannot tell you.

The table is maintained rather than superseded: the slug carries no date, and rows are corrected in place and re-dated when a vendor page changes. Three changes are already queued for the next refresh if they land — Cursor's .cursorrules deprecation moving from future tense to fact, the .windsurf/ fallback paths being dropped, and Junie's legacy guidelines.md losing support. None has a published date, so none is announced here. If you want a second set of eyes on your own instruction estate across several tools, that kind of mapping is part of what our AI transformation engagements cover.