interface-copy
Skill1 starsChanged 52 days agoApache-2.0View on GitHub
Use when writing or reviewing any user-facing words — button labels, errors, empty states, dialogs, tooltips, form hints, or accessible names.
In this repository45 other files it gives its agents.
- docs/agents.mdAGENTS.md
- 3d-game-assetsSkill
- accessibility-evidenceSkill
- accessible-componentsSkill
- agent-memorySkill
- agent-orchestrationSkill
- code-qualitySkill
- code-reviewSkill
- colour-systemsSkill
- component-architectureSkill
- design-judgmentSkill
- design-reviewSkill
- design-tokensSkill
- engineering-disciplineSkill
- information-architectureSkill
- interaction-designSkill
- interface-statesSkill
- layout-compositionSkill
- media-driven-motionSkill
- methodology-feedbackSkill
- micro-interactionsSkill
- mobile-performanceSkill
- motion-designSkill
- motion-physicsSkill
- multiplayer-game-publishingSkill
- platform-androidSkill
- platform-appleSkill
- platform-webSkill
- procedural-surfacesSkill
- public-api-integrationSkill
- rendering-performanceSkill
- responsive-architectureSkill
- reverse-engineeringSkill
- scroll-experiencesSkill
- seo-and-metadataSkill
- ship-readinessSkill
- skill-discoverySkill
- surface-and-depthSkill
- test-driven-developmentSkill
- theming-systemsSkill
- typographic-systemsSkill
- ui-generation-workflowSkill
- vishwakarma-studiosSkill
- visual-feedback-loopSkill
- webgl-experiencesSkill
The file
We split this skill into the rules it gives an agent
8 rules. About 0 are common, 7 are distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈7
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 17
Every string answers a question the user is already asking.
- Distinctive: line 36
Default for product UI: second person, present tense, active voice.
- Distinctive: line 65
Never blame the user: "You entered an invalid date" and "Choose a date on or after today" say the same thing, and only one is an accusation.
- Distinctive: line 76
Use three parts in order: what belongs here, why it is useful, one action.
- Very specific: line 142
Search for lorem, ipsum, TODO, asdf, Oops, and !
- Distinctive: line 527
Run these checks against your own output.
- Distinctive: line 534
If the user read only the buttons and not the body text, could they choose correctly?
- Distinctive: line 557
Read the list of every link name on the page alone — is each one meaningful?
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
<!--
Generated by Vishwakarma. Do not edit this file directly.
Edit the source skill and run `vishwakarma sync` to regenerate.
-->Interface Copy
Words are the only part of an interface the user reads literally; everything else is inferred. A wrong word therefore does more damage than a wrong margin, and faster: nobody has to squint to notice that "Submit" says nothing about what will happen.
Every string answers a question the user is already asking. A button answers "what happens if I press this?" An error answers "what broke and what do I do now?" An empty state answers "is this broken, or have I not started?" Copy that does not answer its question is decoration, and decoration made of text is read before it is found worthless.
1. Voice is constant, register moves
The product should sound recognisably itself everywhere, but register must change — expansive on a landing page, terse and factual in a disk-quota warning. Register is set by how much attention the user has and how much is at stake.
Low stakes — a toast, a menu item — want the fewest possible words. High stakes — deleting a workspace, entering card details — want precision and a slower rhythm, and this is where playfulness becomes offensive. Marketing voice in a settings screen is the commonest register failure: a user three levels deep in billing is working, not browsing, and "Let's get you set up!" reads as a colleague interrupting.
Default for product UI: second person, present tense, active voice.
2. Buttons name the outcome
A button label should be the verb phrase the user would use to describe what they just did. "Submit" describes what the form does to the server; "Create account", "Send invite", "Delete 4 files" describe what the user gets.
This matters most in dialogs, where the eye reaches the buttons before the prose. Buttons reading "OK" and "Cancel" force a re-read to work out which one destroys data — and "Cancel" is catastrophically ambiguous when the action is itself a cancellation ("Cancel subscription?" / "Cancel"). Label the buttons with outcomes and the prose becomes optional: "Keep subscription" and "Cancel subscription" need no explanation. Match the heading's verb: if the dialog asks "Discard changes?", the button says "Discard", not "Yes".
3. Errors: cause, then next action
A useful error has three parts and most shipped errors have none: what happened, why, and what to do next. The user needs neither the stack trace nor an apology. "Something went wrong" fails because it is true of every possible failure; "Invalid input" fails because it names a verdict rather than a cause.
- Bad:
Error: Invalid email.Good:This email address is missing an @ symbol. - Bad:
Oops! Something went wrong.Good:We couldn't save your changes — the server didn't respond. Your draft is stored locally, so try again in a moment.
Never blame the user: "You entered an invalid date" and "Choose a date on or after today" say the same thing, and only one is an accusation. Reserve apology for genuine service failures, so it still means something when used.
4. Empty states are the first lesson
The empty state is seen by every new user and is usually the least-considered screen in the product. "No items" wastes the best teaching moment available.
Use three parts in order: what belongs here, why it is useful, one action. "No saved views yet" plus "Saved views keep a filter and sort order so you can return in one click" plus a single "Create saved view" button turns a dead end into onboarding. One action only; four equal options reproduce the paralysis the state exists to resolve.
Distinguish the three empties that look identical: never had data, filtered to nothing, and failed to load. The second needs "Clear filters", not "Create your first item"; the third needs "Retry", and rendering it as an empty list tells the user their data is gone.
5. Confirmations state consequences
"Are you sure?" moves the burden of understanding onto the user at the moment they are least able to carry it. It tests confidence, not comprehension, and is dismissed reflexively.
State the irreversible part instead: "Delete 12 files permanently? They can't be recovered." Name the object and the count — "Delete project" is weaker than "Delete Acme Redesign and its 40 tasks" — and say when something is reversible, because most confirmation anxiety is uncertainty about reversibility. Better still, for reversible actions drop the dialog and offer undo; a confirmation taxes the common case to guard against the rare one.
6. Tone failures with names
Exclamation marks perform an enthusiasm the reader does not feel, and in a failure message they are hostile: "Something went wrong!" is cheerful about your data loss. "Oops", "Whoops", "Uh-oh" signal that the system finds its own failure charming. Mock-cheerful failure ("our hamsters need a break") carries no diagnostic content and cannot be searched for in a support forum. Jargon leaking upward — "null reference", "422", "token expired" — is implementation vocabulary; translate it: "Your session ended. Sign in to continue."
7. Settled mechanics
Case, numerals, time, plurals, truncation, placeholders and translation headroom are decided
once for a codebase and then applied without further judgment. copy-mechanics.md has each
with its reasoning: sentence case over title case, Intl.PluralRules over an appended s,
relative time only while recency is the point, a visible <label> rather than a placeholder,
and roughly 30 per cent more room than the English string needs.
8. The accessible name is copy
Screen reader and voice control users hear the accessible name, not the pixels. For a button
it resolves from aria-labelledby, then aria-label, then text content — and aria-label
silently overrides visible text.
That override is where copy breaks accessibility. A button reading "Save" with
aria-label="Submit form" cannot be operated by a voice user saying "click Save", because
the utterance is matched against the accessible name. WCAG 2.2 SC 2.5.3 (Label in Name)
requires the accessible name to contain the visible label: a name may extend the visible
text but must never replace it.
Link text must stand alone, because screen reader users navigate by listing every link with its context stripped. Eleven "Read more" links produce eleven identical entries. Write the destination in: "Read the migration guide".
9. Before shipping
Search for lorem, ipsum, TODO, asdf, Oops, and ! in message strings. Placeholder
copy reaching production is not a rare accident — it occupies the right shape, so it survives
visual review and is caught only by a string search.
Rules
MUST NOT — Do not use exclamation marks, "Oops", "Whoops", or jokes in messages that report a failure.
Why: Failure copy is read by a user who has just lost time or data. Performed cheerfulness signals that the system does not consider the loss serious, which converts irritation into distrust, and the words displace the diagnostic information the message should have carried.
Incorrect:
Oops! Something went wrong!Correct:
We couldn't reach the server. Your draft is saved locally.MUST NOT — Do not use "Click here", "Read more", "Learn more", or "This link" as the entire accessible name of a link.
Why: Screen reader users routinely navigate by generating a list of every link on the page with all surrounding context stripped, so identical generic names produce a list of indistinguishable entries and the page becomes unnavigable by that route.
Source: WCAG 2.2 SC 2.4.4 Link Purpose (In Context)
Exceptions:
- The visible text may remain "Read more" if an aria-label preserves it and extends it, e.g. "Read more about pricing".
MUST NOT — Do not ship lorem ipsum, "TODO", dummy names, or invented testimonial text in any user-facing string.
Why: Placeholder prose reads as intentional in review because it occupies the correct shape, so it passes visual inspection and is caught only by a string search. Invented attributed quotes are additionally a misrepresentation, not merely an omission.
MUST — Label every button with the verb phrase describing the outcome, not with a generic acknowledgement such as Submit, OK, Yes, or Continue.
Why: Users read buttons before body text, so the label is often the only text consulted before committing. A generic label forces a return to the prose to determine consequence, and under time pressure that re-read does not happen.
Exceptions:
- Platform-standard dialogs whose button labels are supplied by the operating system.
Incorrect:
<button type="submit">Submit</button>Correct:
<button type="submit">Create account</button>MUST — A confirmation dialog must state what will happen and whether it is reversible, rather than asking whether the user is sure.
Why: Asking for certainty tests confidence rather than comprehension and adds no information the user did not already have, so it is dismissed reflexively. Naming the object, the scope, and the reversibility supplies the fact the user is missing.
Exceptions:
- Reversible actions, where an undo affordance should replace the dialog entirely.
Incorrect:
Are you sure? [Yes] [No]Correct:
Delete Acme Redesign and its 40 tasks? This cannot be undone. [Delete project] [Keep project]MUST — Every error message must name the specific cause and the next action the user can take.
Why: An error is an interruption the user did not choose; its only value is restoring forward motion. A message true of every failure carries no information and converts a recoverable state into a support request.
Source: WCAG 2.2 SC 3.3.1 Error Identification and SC 3.3.3 Error Suggestion
Exceptions:
- Security-sensitive failures where naming the cause would leak information, such as distinguishing an unknown account from a wrong password.
Incorrect:
Something went wrong. Please try again later.Correct:
report.pdf is 24 MB and the limit is 10 MB. Compress it or upload a smaller file.MUST — Distinguish never-had-data, filtered-to-nothing, and failed-to-load states with different copy and different actions.
Why: The three states share a visual shape but require opposite responses: create something, clear the filter, or retry. Rendering a load failure as an empty list actively misinforms the user that their data no longer exists.
MUST — Every form field must have a persistent visible label; placeholder text may only supplement it with a format example.
Why: Placeholder text vanishes on focus, so the field name disappears at the moment it is needed, which particularly harms users with memory or attention differences. Placeholders also typically fail contrast requirements and are not a reliable accessible name.
Source: WCAG 2.2 SC 3.3.2 Labels or Instructions
Incorrect:
<input type="email" placeholder="Email address" />Correct:
<label for="email">Work email</label>
<input id="email" type="email" placeholder="name@company.com" />MUST — When a control has visible text, its accessible name must contain that text, in the same order — never replace it with an unrelated aria-label.
Why: Speech input matches a spoken command against the accessible name rather than the rendered pixels, so an aria-label that discards the visible word makes the control unreachable by voice and gives the user no feedback about why.
Source: WCAG 2.2 SC 2.5.3 Label in Name
Incorrect:
<button aria-label="Submit form">Save</button>Correct:
<button aria-label="Save draft to your workspace">Save draft</button>MUST — Resolve plurals with Intl.PluralRules or ICU message syntax, never by appending an "s" or shipping "item(s)".
Why: English has two plural categories but the CLDR set spans six, and languages such as Polish, Russian, and Arabic select a form based on the numeral itself. String concatenation therefore produces grammatically wrong output in most target locales, and "(s)" is unreadable aloud.
Source: ECMA-402 Intl.PluralRules; Unicode CLDR plural categories
Incorrect:
const label = `${count} item${count === 1 ? "" : "s"}`Correct:
const label = t("itemCount", { count }) // "{count, plural, =0 {No items} one {# item} other {# items}}"MUST — Never size a container to fit its English string; allow roughly 30% expansion for running text and more for short labels, and use logical properties for direction.
Why: German running text averages around 30% longer than English and short UI labels can more than double, so a container fitted to English clips or wraps badly on translation. Physical properties such as padding-left do not mirror in right-to-left locales, where icon order, progress direction, and arrows must all flip.
Source: W3C Internationalisation guidance on text expansion
Incorrect:
.button { width: 88px; padding-left: 12px; text-align: left; }Correct:
.button { min-width: 88px; padding-inline: 12px; text-align: start; }SHOULD NOT — Do not phrase validation messages as accusations directed at the user ("You entered an invalid…"); state the constraint instead.
Why: Second-person blame adds an emotional cost without adding information, and the constraint phrasing is strictly more useful because it tells the user what a valid value looks like rather than only that theirs was not.
Incorrect:
You entered an invalid date.Correct:
Choose a date on or after today.SHOULD NOT — Do not use marketing or promotional voice in utility surfaces such as settings, billing, errors, or destructive dialogs.
Why: Register signals what kind of interaction is happening. A user deep in billing is performing a task with consequences, and persuasive enthusiasm there reads as an interruption by something that wants their attention rather than a tool that is helping them finish.
Incorrect:
Ready to supercharge your workflow? Let's get you set up!Correct:
Your plan renews on 12 August 2026.SHOULD NOT — Do not open with a category inversion — "It's not X, it's Y", "X, reimagined", "X, but for Y", "Stop doing X. Start doing Y." — unless the same block names the specific capability that makes the new category true.
Why: The construction asserts a reclassification while supplying nothing a reader can check, and it spends the most-read line in the product describing a category the product is not in. Its frequency in generated marketing copy is now high enough that the shape itself is read as authorship rather than as a claim, so it discredits whatever follows it.
Exceptions:
- Copy correcting a misclassification the audience has demonstrably already made, where naming the wrong category is the information: "This is a compiler, not a linter — it rejects programs a linter would only warn about."
- Comparative copy where X is a named competitor or a prior version and the difference is stated in measurable terms rather than implied by the inversion.
Incorrect:
It's not a note-taking app, it's a second brain.Correct:
Every note is linked to the meeting, person, and project it came from.SHOULD NOT — Do not prefix feature-list items with checkmarks or tick icons; reserve the tick for state the system actually verified, such as a completed step, a passing check, or a column in a two-sided comparison table.
Why: A tick is a verification glyph everywhere else in an interface, so on an unchecked feature list it asserts a confirmation that nothing performed. It also gives every item identical weight, removing the ranking the reader came to the list for, and it reproduces the left half of a feature matrix, which invites the reader to notice that the competitor column is missing.
Exceptions:
- Checklists, onboarding progress, and status lists, where the tick reports a fact the product confirmed.
- A real comparison table where the tick contrasts with a cross and both products are named.
Incorrect:
✓ Fast ✓ Secure ✓ Scalable ✓ Easy to useCorrect:
Cold starts under 50 ms in 18 regions. SOC 2 Type II with keys you rotate yourself. Tested to 40,000 requests per second.SHOULD NOT — Do not use three-comparative cadences such as "faster, simpler, smarter", or more than one standalone sentence fragment per section, in place of a measured claim.
Why: Three is the shortest list that sounds complete, so the slot gets filled to length and the third member is chosen for stress rather than meaning. None of the comparatives names a baseline, which makes all three unfalsifiable while the rhythm supplies the impression that a claim was made. Fragment emphasis works the same way, lending a phrase the weight of a full stop without a finite verb to carry a claim.
Exceptions:
- A triad whose three members are each measured or independently verifiable.
- A single fragment answering an immediately preceding question, or a tagline the reader encounters once rather than as a rhythm.
Incorrect:
Faster, simpler, smarter. Built for teams. No setup. Just results.Correct:
Builds finish in 40 seconds instead of six minutes, and two people can edit the same file at once.SHOULD — Give every empty state three parts: what belongs here, why it is useful, and exactly one action.
Why: The empty state is seen by every new user before any populated screen, making it the earliest available teaching surface. A bare "No items" cannot be distinguished from a fault, and multiple competing actions reproduce the paralysis the state exists to resolve.
SHOULD — Choose sentence case or title case once and apply it to every button, heading, tab, and menu item in the product.
Why: Title case has no agreed rule set across English variants for prepositions, conjunctions, and hyphenated words, so a codebase using it drifts into inconsistency that readers perceive as carelessness without being able to name it. Sentence case has one rule and survives contributor turnover.
SHOULD — Write the zero case as its own string rather than letting it fall through to the plural form.
Why: Zero is semantically a different message from a count: it reports absence rather than quantity. "No unread messages" answers the user question, whereas "0 unread messages" makes the reader parse a numeral to reach the same conclusion.
SHOULD — Use relative time only for recent events where precision does not matter, and always expose the absolute timestamp via a <time datetime> element or tooltip.
Why: Relative time is easier to read for recency but loses resolution as it ages and cannot be compared, cited, or reconciled with an external record — which is exactly what a user needs from an audit log, receipt, or scheduled event.
Incorrect:
<span>2 years ago</span>Correct:
<time datetime="2024-03-04T09:12:33Z" title="4 March 2024, 09:12 UTC">4 March 2024</time>SHOULD — When truncating a value the user may need to act on, keep the full value reachable on hover, focus, or in a detail view, and truncate filenames in the middle so the extension survives.
Why: Truncation discards information the layout could not fit, not information the user did not need. Without a recovery path the user cannot distinguish two similarly-prefixed items, and a trailing ellipsis on a filename removes the extension, which is often the most identifying part.
SHOULD — Replace implementation vocabulary — status codes, exception names, internal entity names — with words from the user’s model of the product.
Why: Implementation terms are precise only for people who can act on them. For everyone else they add reading cost and imply the failure is theirs to diagnose, which stalls the user at exactly the point the message should be moving them forward.
Exceptions:
- Developer-facing tools, and error surfaces where a copyable code genuinely helps support — include it as secondary detail, not as the message.
SHOULD — Keep em dashes to at most one per short paragraph and two or three per page of marketing copy; above that rate, re-punctuate each one as the comma, colon, semicolon, or full stop the relation actually calls for.
Why: The em dash accepts every relation between two clauses, so it lets a writer join two thoughts without deciding whether the relation is apposition, consequence, contrast, or a sentence break. A high rate is therefore evidence those decisions were skipped, and because punctuation is what varies the rhythm of prose, one mark carrying every join gives every sentence the same shape.
Exceptions:
- Quoted speech, transcripts, interviews, and fiction, where the dash reproduces how someone talks.
- A matched pair of dashes bracketing a genuine parenthetical, which is one construction rather than two connectors.
- Dense reference prose where commas are already carrying a list inside the clause and a comma would be ambiguous.
Incorrect:
Deploy in seconds — no config, no YAML — and scale automatically — from one user to a million.Correct:
Deploy in seconds. No config files, no YAML. Traffic scales from one user to a million without a change on your side.SHOULD — Let bullet length and grammatical shape follow what each item is worth, and put the most consequential item first, rather than trimming every item to the same three-word noun phrase.
Why: A reader uses relative length and position as a ranking signal before reading any of the words, so a list where every item is the same size states that nothing in it matters more than anything else. Uniformity of that kind comes from filling a template to a round number, and the items added to reach the number are the ones with nothing behind them.
Exceptions:
- Specification, parameter, and option lists, where parallel construction is what lets the reader compare values in the same dimension.
- Localised sets where parallel structure is a translation or legal-review requirement.
Incorrect:
Fast builds / Simple setup / Great support / Full control / Team readyCorrect:
Builds finish in 40 seconds, down from six minutes.
Setup is one command; there is no config file.
Support answers in under 4 hours, including weekends.SHOULD — Name each pricing tier after the buyer or the limit it fits, and state every difference between adjacent tiers as a number, so a reader can self-select without contacting anyone.
Why: The tier name is the reader’s fastest route to "which one is me" and is read before any bullet list, so a name from an internal ladder such as Starter/Pro/Enterprise answers that question for nobody and forces all three lists to be read. A boundary the copy states as "advanced features" rather than a number is usually a boundary that was never decided, and a "Most popular" badge substitutes a social cue for the fact the reader was looking for.
Exceptions:
- Contracts genuinely priced per negotiation, where "Contact us" is accurate; state the threshold at which a reader should make contact, such as above 50 editors or when data residency is required.
- A single usage-based rate, where there are no tiers to select between.
Incorrect:
Starter: Free. Pro: $29/mo (Most popular) — everything in Starter, plus advanced features. Enterprise: Contact us.Correct:
Solo: $0. One editor, 3 projects, 7-day history.
Team: $29 per editor per month. Unlimited projects, SSO, audit log. For 2 to 50 editors.
Company: from $18,000 a year. Adds SAML, data residency, 99.9% uptime. Contact us above 50 editors.SHOULD — A subtitle must add a fact its heading does not already contain; if striking the shared words leaves no new claim, delete the subtitle and let the heading run one line longer.
Why: The subtitle slot exists in the layout before anyone has decided what belongs in it, so it gets filled by paraphrase. It is the second most-read line on the page, and a reader who finds no new information there learns the page repeats itself, after which everything below is skimmed rather than read.
Incorrect:
Ship faster with automated deploys / Deploy automatically and ship your product fasterCorrect:
Ship faster with automated deploys / Push to main and the new version is live in about 90 seconds.Before reporting completion
Run these checks against your own output. Answer each question explicitly rather than assuming the answer, because the point of the exercise is to notice what you did not notice while building.
Confirm labels and headings answer the user’s question. (blocking)
- For every button on screen, does the label name the outcome rather than the mechanism?
- If the user read only the buttons and not the body text, could they choose correctly?
- Is the case convention (sentence or title) identical across every label, heading, tab, and menu item?
- Does any string use marketing register on a utility surface?
Confirm every failure message is actionable. (blocking)
- List every error string. Does each name a specific cause rather than a generic failure?
- Does each state what the user should do next?
- Does any error contain an exclamation mark, "Oops", a joke, or an unexplained status code?
- Does any message blame the user rather than state the constraint?
Confirm empty states teach rather than terminate.
- Does the empty state say what belongs here and why it is useful?
- Does it offer exactly one action?
- Are the never-had-data, filtered-to-nothing, and failed-to-load cases distinguished by different copy and different actions?
Confirm spoken names match visible words. (blocking)
- Does any aria-label replace rather than extend the visible text of its control?
- Does every icon-only control have a name describing the action, not the icon?
- Are any two controls on the screen announced with the same name?
- Does every form field have a persistent visible label rather than a placeholder acting as one?
- Read the list of every link name on the page alone — is each one meaningful?
Confirm the layout survives translation and direction change.
- Does every layout hold with each string 40% longer?
- Is any container sized to the exact width of its English label?
- Are directional styles written with logical properties (padding-inline, text-align: start) rather than left and right?
- Are counts, dates, currencies, and units formatted through Intl rather than concatenated?
Fail if placeholder or draft copy reaches user-facing strings. (blocking)
grep -rniE "lorem ipsum|dolor sit amet|asdf|\\bTODO\\b|FIXME|Oops|Whoops" --include="*.tsx" --include="*.jsx" --include="*.vue" --include="*.svelte" --include="*.json" src && exit 1 || exit 0Further reference
These are not loaded by default. Read one only when its question is the question you currently have.
references/copy-mechanics.md— Sentence case or title case, relative or absolute time, how to pluralise correctly, where to truncate, and how much room translation needs?references/copy-rewrites.md— I have a specific bad string — a button, error, empty state, dialog, tooltip, or notification — what does the corrected version look like and why?references/accessible-naming.md— How do I write aria-label, alt text, and link text so screen reader and voice control users get the same information as sighted users?references/generated-copy-tells.md— This marketing or landing-page copy reads as AI-generated but I cannot say why — which construction is doing it, what is the mechanism, and what does the rewrite look like?
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

