Skip to content
Pauta6.1.0

Coding agents and LLMs

Pauta has two readers: the developer, and the coding agent working beside them. The pages are written for the developer. Every edition also serves the same rules as plain text and as JSON, so an agent (Claude Code, Codex, Copilot, or any LLM that can fetch a URL) reads the exact rules of the edition the app is built to, instead of what it remembers of some other one.

FileWhat it holdsRead it when
llms.txtThe index, in the llms.txt format: what Pauta is, how to start an app, build a feature and audit an app, and every page with its rule count and prefix.First. It is short.
llms-full.txtEvery rule of the edition as Markdown, grouped by page and section, with every link made absolute. No prose, screenshots or mockups.Building or auditing: the whole rule set in one read.
rules.jsonEvery rule as data, with the levels, the three kinds of check and the audit statuses explained, every page, and every retired ID.Scripts, filtering by page or level, comparing editions.
<page>.mdOne page as Markdown: its prose, rules and spec tables, each mockup as a sentence and its numbered parts as a list. keyboard.md beside keyboard/, templates/about.md beside templates/about/.Building a feature: the page it belongs to, in one small read.
changelog.mdThe Changelog page as Markdown: how editions are numbered and what each one changed.Moving an app to a newer edition.
skills/pauta-audit/SKILL.mdThe audit procedure as an agent skill: how to judge each rule, read pauta.json and write the report.Auditing an app.
audit.schema.json, rules.schema.jsonThe JSON Schema of an audit report and of rules.json.Writing or checking a report; reading rules.json in code.
The page itselfThe prose, screenshots, mockups and spec tables around the rules.When the picture matters.

Each one lives under its edition: https://pauta-project.dev/4.3.1/llms.txt, .../4.3.1/rules.json, .../4.3.1/keyboard.md. latest/ always holds the newest edition, and the site root holds a copy of the latest llms.txt, llms-full.txt and rules.json, where agents look first. Every link inside these files points at the numbered edition, so a rule an agent cites keeps its address. llms.txt, llms-full.txt and rules.json are served from 2.0.0 on, the page Markdown, the skill and the schemas from 4.3.1, and changelog.md from the edition after 4.5.0; the 1.x editions, from before rules had IDs, have their pages only. versions.json at the site root lists every published edition.

The single Markdown file is the whole guide in one file. It is large. Its pages and rules are the site’s, and each mockup reads as it does in <page>.md: a sentence saying what it shows.

Every rule is one testable statement with a permanent ID. In llms-full.txt it is the same list item as on its page: the ID and level in bold, the statement, then Why, Check, sometimes Command, and Reference lines. In rules.json the same rule is:

{
"id": "KBD-07",
"level": "MUST",
"changed": "x.y.z",
"page": "keyboard",
"pageTitle": "Keyboard first",
"section": "Command Palette",
"appliesTo": null,
"url": "https://pauta-project.dev/<edition>/keyboard/#kbd-07",
"text": "The palette and every picker (a card that lists choices to filter and pick one from: [QOPEN-01](https://pauta-project.dev/<edition>/templates/quick-open/#qopen-01)) are in-scene overlays in the window's root `StackPane`, never a `javafx.stage.Popup`.",
"why": "on Windows a Popup does not reliably take OS keyboard focus, and the whole app can stop receiving keystrokes.",
"check": {
"kind": "code",
"text": "the palette and every picker add their node to the root `StackPane`; none creates a `Popup`.",
"command": null
},
"reference": "Editora [`ui/OverlayHost.java`](https://github.com/adriandeleon/Editora/blob/f314ca3b117829d25dbaa63be6ba2353fd898fd8/src/main/java/com/editora/ui/OverlayHost.java), the one host installed in the scene's root `StackPane`, which shows the palette ([`ui/CommandPalette.java`](https://github.com/adriandeleon/Editora/blob/f314ca3b117829d25dbaa63be6ba2353fd898fd8/src/main/java/com/editora/ui/CommandPalette.java)), the picker ([`ui/QuickOpen.java`](https://github.com/adriandeleon/Editora/blob/f314ca3b117829d25dbaa63be6ba2353fd898fd8/src/main/java/com/editora/ui/QuickOpen.java)) and the picker's relatives ([QOPEN-12](https://pauta-project.dev/<edition>/templates/quick-open/#qopen-12))."
}
  • id is permanent. It is never renumbered or reused; a removed rule leaves a gap. The prefix names the page (KBD is Keyboard first), and llms.txt lists every page’s prefix. rules.json ends with retired, every ID no longer in use: the edition that removed it and see, the rule that holds its requirement now (with its seeUrl), or null. A pauta.json entry for a retired ID is stale; move it to the see rule, if the app still misses it.
  • level is MUST (a release blocker for a new app), SHOULD (expected unless the app has a stated reason) or MAY (worth copying when the app grows into it).
  • check.kind says how the rule is verified: code by reading or searching the source, test by a named test the build runs, by hand by a person, in the running app or where the check sends them (DOCS-16 reads the repository’s pull requests), which source alone can’t settle. A rule has a test check only when its statement requires the test (“A test holds it.”, or a rule that is itself about a test). A code check may end “A test that asserts it is the surest evidence”: that test is evidence, and an app without it still meets the rule by the behavior its code shows. A check that reads what the app’s checkout does not hold, the website’s repository or a repository setting, says so in its last sentence (not-checked).
  • check.command, on some code checks, is one POSIX shell command (Linux, macOS, Git Bash or WSL): run it with sh -c from the app’s repository root. Exit 0: the check passes; still read the rest of the rule and cite the file:line that meets it. Any other exit: the check fails, and what it printed is the evidence (OS-02’s lists the classes that read os.name). Every command only reads, and opens with a guard that fails it outside an app’s root. It is null where reading the code takes judgment. A search written inside check.text (the grep of PERF-01, say) is not a command in this sense: it lists candidate lines for a person or an agent to judge, and its exit code is no verdict.
  • reference names the Editora file, or the other app, that does it. The agent copies from there. From the first edition released after 4.5.0, its links point at the commit each app was at when the edition came out, so they show the code as it was then; blob/master in a page’s source becomes that commit in what the edition serves. Editions up to 4.5.0 link the apps’ default branches.
  • Names in a rule. Some names in a rule’s text are required as spelled, because a user, the docs or another tool finds the thing by that name: a command id, a catalog key, a chord, a command-line flag, an environment variable, a property in pom.xml or pauta.json, and a file the repository must carry, whether a document at its root or a resource or tool configuration the rule names. For example help.about, palette.show, --config-dir, pauta.version, AGENTS.md, emacs.json, messages_<code>.properties and jreleaser.yml. A platform API the rule names (javafx.stage.Popup, SystemTray, Platform.setImplicitExit(false), Future.get()) is the requirement itself: the JDK and JavaFX spell it one way for every app. Any other name is how Editora spells it, and the app’s equivalent meets the rule: a class, method or constant of the app’s own code, a CSS file or style class, a branch or workflow file, a file the app writes as it runs (its config files).
  • why is null when the rule has no reason worth stating.
  • appliesTo is null for a rule that covers every app, unless its own text says otherwise. Otherwise it is the Applies to line its page or section opens with: “an app whose work goes on without its window” for the TRAY rules, and “every app that sends desktop notifications…” for TRAY-09 to TRAY-11. pages lists every page with its prefix, its Markdown address and its appliesTo.
  • changed is the edition in which the rule’s level, scope or statement last changed, as a version ("5.0.0"). A checked entry in an app’s pauta.json whose edition is older than it is stale (PAUTA-18). A rule unchanged since it was added has the edition that added it, 2.0.0 at the earliest; a rule’s Why, Check, Command and Reference do not move it. Editions up to 5.0.0 don’t have the field. In next/, the batch that is not an edition yet, a rule new or changed since the last edition has "next".
  • statuses are what an audit reports for a rule (Audit an app), and audit is the address of the report’s schema.

An app is built to one edition, the one pauta.version in its pom.xml names (see About window). Its agent should read that edition, not latest/: a newer edition may hold MUSTs the app doesn’t meet yet, and an agent that reads them will “fix” what nobody asked it to.

Say so in a section of the app’s AGENTS.md (PAUTA-17). The app template’s AGENTS.md has one, headed “Pauta”; an app made before it adds this:

## Pauta
This app is built to Pauta <edition>. Its rules: https://pauta-project.dev/<edition>/llms.txt
(every rule in one file: https://pauta-project.dev/<edition>/llms-full.txt).
- Before building a feature, read the rules on its page and copy from the code each rule's Reference names.
- Cite a rule by its ID and edition: "KBD-07 (Pauta <edition>)".
- Read the numbered edition, never latest/. Moving to a newer edition is its own change: pauta.version in
pom.xml, this file and the README change together.

Use the same number everywhere the app names its edition: pom.xml, AGENTS.md, the README and the About window. An app whose conformance is Partial writes is working toward Pauta <edition> instead of is built to, as its README and About do (Conformance).

Give the agent the rules of the page the feature belongs to, the page itself when it has a mockup or a spec table, and the reference code. For example:

Add a find bar to the editor. Read the Find bar rules (FIND-*) in
https://pauta-project.dev/<edition>/llms-full.txt and the page
https://pauta-project.dev/<edition>/templates/find-bar/ for its layout. Copy from the
Editora files each rule's Reference names. Meet every MUST; list the SHOULDs you
skipped and why.
  • Template pages keep their mockups and spec tables on the page; llms-full.txt holds only their rules. Give the agent the page’s address too.
  • Reference code is in Editora, or the app the reference names. An agent with Editora checked out beside the app reads it faster and more reliably than over the web.
  • When the rule and Editora disagree, the code is right until someone checks. Ask the agent to report the difference instead of picking a side, and fix the guide or the app in the same change.

The procedure is the pauta-audit skill, which every edition serves at skills/pauta-audit/SKILL.md. Copy it into the app’s .claude/skills/pauta-audit/ (the app template ships it) or point the agent at it. Use the copy of the edition the app is built to (the newest edition’s for an app on an edition before 4.3.1, which serves none), and replace the app’s copy when the app moves to a newer edition (PAUTA-11). The skill decides how to audit and how to write the report, and rules.json supplies the rules. llms.txt and llms-full.txt carry the procedure in short; where an edition’s short version differs from the skill, the skill wins:

  1. Find the edition the app implements: pauta.version in its pom.xml, or failing that the “Built to Pauta” line of its README. Read that edition’s rules. If the app names none, ask which edition to audit against and offer the newest; with no one to ask, use the newest and say so in the report. A 1.x edition has no rules.json: audit against 2.0.0 or newer, and report what the app’s own edition didn’t ask for as upgrade work.
  2. Take each rule in turn. A code check: run its check.command when it has one, else read or search the source as it says. A test check: find the named test, or the app’s equivalent, and confirm the build runs it. A by hand check: the source can’t confirm it, so report it as manual, unless pauta.json lists the rule (step 4) or the source plainly shows it unmet, which is a fail like any other. A check that needs the website’s repository or a repository setting, which the app’s checkout does not hold, is not-checked when the audit can’t reach it. Which names in a rule are required as spelled is under What a rule looks like.
  3. A rule applies to every app unless it is limited: its text names the apps it covers (“An app with user documents…”), its page or section opens with an Applies to line, or it details a limited rule (PAUTA-05). An app outside a limited rule gets limited, and needs no exception. Those three are the only places a scope is written: an audit never infers one from what the app lacks.
  4. Read the app’s pauta.json (Conformance). An unmet rule it lists gets the entry’s kind as its status; MAYs need no entry. An unmet MUST or SHOULD it doesn’t list is fail; so is an unmet MAY, which has no consequence. An entry for a rule that passes or is limited, and a deviation or pending entry with no link, are problems. A by-hand rule it lists as an exception gets that kind too, not manual. Its checked list records the by-hand rules a person checked in the running app and found met (PAUTA-18): a manual result carries its entry, and one with no entry is reported as not recorded, which is not a failure: the list is the app’s to keep or not, and PAUTA-18 governs only the entries that are there. An entry older than the rule’s changed is stale, and a problem.
  5. Report one result per rule: its ID, level, status and evidence as file:line. A failed MUST blocks a release; a manual or not-checked MUST does not, and stays in the report until someone settles it.

Every report uses the same statuses, which rules.json also lists:

StatusMeansNeeds an exception?
passMet.No; an entry for it is stale.
failNot met, and pauta.json doesn’t list it. A failed MUST blocks a release; a failed MAY has no consequence.A MUST or SHOULD, yes: this is what the audit is for. A MAY, no.
manualChecked by hand: the source leaves it open. A person checks it, in the running app or where the Check sends them (the repository’s pull requests, for DOCS-16), and may record it in pauta.json’s checked list once it is met; the report gives that entry, or says “not recorded”.No, recorded or not. Found unmet by the person who checks, it gets one, and its status is that kind from then on.
not-checkedThe Check needs something the audit could not read: another repository (the website’s), what the repository’s host keeps and no file shows (a repository setting, an issue’s state), or the rules of the edition the app names, in an audit run against another (ABOUT-07). Not a finding and not a pass; the evidence says what would settle it. It does not block a release.No. Found unmet by whoever settles it, it gets one.
pending, deviation, not-applicable, app-onlyNot met, and pauta.json lists it as that kind.It has one.
limitedThe rule covers apps with a feature this one lacks.No; an entry for it is wrong.

The skill writes the report twice: docs/pauta-audit.json, which follows the edition’s audit.schema.json so two audits can be compared by a script, and docs/pauta-audit.md, the same results as tables for people. A prompt as short as this one is enough:

Audit this app against Pauta <edition> with the pauta-audit skill
(https://pauta-project.dev/<edition>/skills/pauta-audit/SKILL.md). Failed MUSTs first.

Each result reads like this:

KBD-07 MUST pass src/main/java/com/example/app/ui/CommandPalette.java:57
TBAR-01 MUST fail src/main/java/com/example/app/ui/MainToolbar.java:88 builds an icon button by hand
SET-02 MUST manual the Settings window's layout needs the running app; checked 2026-11-02 against <edition>
PWIN-02 MUST manual compare the Pauta window with pauta.json in the running app; not recorded
TRAY-01 MUST limited no tray: no SystemTray or TrayIcon in src/
PAUTA-13 MUST not-checked needs the website's repository: its docs have a Pauta conformance page
DIAG-07 SHOULD not-applicable pauta.json: the app depends on no outside tool, so a Doctor has nothing to check
MENU-15 MUST pending pauta.json: issue #212
  • Split a large audit. An edition has hundreds of rules. Give each page, or each prefix, its own pass or its own agent, and start with the MUSTs (jq '.rules[] | select(.level == "MUST")' rules.json). Each page’s Markdown (<page>.md) is a small read for one pass.
  • Check the evidence. A verdict without file:line is a guess. Open a sample of the passes as well as the failures before you act on the report.
  • Manual means a person. A by hand check needs the running app, often on a particular OS, or in a few rules a look at the repository’s pull requests (DOCS-16). The agent lists them; someone runs them and may record each one met in pauta.json’s checked list (PAUTA-18), which the next audit reads back. One with no entry is reported as not recorded: that is not a failure, it needs no exception, and it leaves the conformance as it is. One the person finds unmet gets an exception like any other rule (PAUTA-04). The agent settles one case by itself: a by-hand rule the source plainly shows unmet is a fail, with the file:line.
  • Not checked means out of reach. A few checks read the app’s website repository (DOCS-01, PAUTA-13) or a repository setting (REL-19). An audit run from the app’s checkout alone reports them as not-checked and says what would settle each; with the website checked out beside the app, or gh signed in, it judges them. The checkout still settles what it holds: an app whose repository and About name no website fails DOCS-01, DOCS-02 and PAUTA-13 outright, with no website to read. And an audit run against an edition other than the one the app names reports ABOUT-07 as not-checked when About says Built to: only the app’s own edition’s rules settle it.
  • Nothing to judge is not a guess. A rule about how a required part behaves (the find bar of a log shown as text, the Pauta window) fails with the rule that requires the part, and its evidence names that rule, so a missing find bar reads as one cause. So does a rule that limits every item of a kind some rule requires: with no keymap file, the rule on a keymap’s macOS variant fails with the rule that requires the files. A rule that constrains things no rule requires an app to have (every entry, row or file of a kind) passes in an app that has none. The skill’s step 4 has the order.
  • Gaps per app on this site is the same audit, done by hand for the six apps. An agent’s report for one of them should agree with Gaps per app; where it doesn’t, one of them is wrong.

The version says how much work the move is (how editions are numbered): a newer patch asks nothing of the app, a newer minor asks it to add, a newer major asks it to undo or redo something. The Changelog says what each edition added and which rules became stronger. To see it rule by rule, compare two editions’ rules.json:

Terminal window
old=2.0.0 new=2.0.2
diff <(curl -s https://pauta-project.dev/$old/rules.json | jq -r '.rules[] | "\(.id) \(.level)"' | sort) \
<(curl -s https://pauta-project.dev/$new/rules.json | jq -r '.rules[] | "\(.id) \(.level)"' | sort)

A line only in the new edition is a new rule or a raised level; one only in the old is a retired rule (the new edition’s retired list says where it went) or a lowered level. A rule whose changed is newer than the app’s edition has a new level, scope or statement since then. Then, in one change (PAUTA-11):

  1. Replace src/test/resources/pauta-rules.json with the new edition’s, and the app’s copy of the audit skill (.claude/skills/pauta-audit/SKILL.md in an app made from the template) with the one the new edition serves.
  2. Remove the checked entries the new rules.json makes stale, those whose edition is older than their rule’s changed, and have a person check those rules again (PAUTA-18).
  3. Give the agent the list and the new edition’s llms-full.txt, and have it meet each new MUST or record it in pauta.json. Run the audit again.
  4. Change pauta.version, AGENTS.md and the README together.

This repository has its own AGENTS.md: the rule syntax, each page’s prefix, and how a change becomes an edition. npm run check fails on a malformed rule, and rules.json, llms.txt and llms-full.txt are built from the pages, so a rule written on a page reaches agents in the next edition with nothing else to update.

Pauta 6.1.0 · · Changelog · This page in the latest edition