Righter

A UX writing skill. Review existing UI copy against a defined set of principles, or write new copy from scratch applying those principles from the start.

Reference files: read these when relevant:

• data/components.json: per-component writing rules (Alert Dialog, Toast, Inline Alert, Helper Text, Alert Banner, Tooltip, Empty State, Permission Prompt). Look up by component id, don't scan the whole file.
• references/accessibility.md: read when reviewing or writing labels, buttons, link text, form errors, or icon-only controls, or whenever accessible-copy might apply
• references/tone.md: read when copy lands on an emotional or high-stakes moment (errors, destructive actions, first use, success)
• references/voice-chart.md: read only when someone asks to define, document, or audit a brand voice
• references/figma.md: read only when someone shares a Figma link and asks for a copy review
• references/email.md: rules for product transactional emails (subject lines, preheaders, body, CTAs, footers, structure)
• data/email-benchmarks.json: numeric deliverability benchmarks (open rate, CTOR, CTR, unsubscribe thresholds) referenced by references/email.md. Query by metric.
• data/weakeners.json: structured word/phrase lists for all weakener categories. Query by category id, don't scan the file top to bottom.
• data/phonaesthetics.json: structured sound concepts and cluster table for word choice
• data/rhetorical-devices.json: literary/rhetorical devices (metaphor, personification, epithet, anaphora, etc.) for copy that's allowed to have personality: marketing surfaces, taglines, empty states, feature names. Not for error messages or system copy.
• references/marketing-copy.md: read when writing or reviewing marketing-site copy (hero, section heads, feature cards, buttons, meta descriptions). Word limits per slot and the mkt-* rules.
• data/marketing-limits.json: word limits per marketing slot, banned puffery, and check severities. Query by slot id.
• scripts/check-slots.mjs: counts words per marketing slot and checks limits, puffery, em dashes, and acronyms: node scripts/check-slots.mjs hero-h1 "copy text". Run it instead of counting by hand.
• scripts/ari.mjs: computes ARI, grade level, and word/character/sentence counts exactly. Run it instead of calculating Reading Metrics by hand when Node is available: node scripts/ari.mjs "copy text" or node scripts/ari.mjs --before "..." --after "...".

Every numbered principle below and every rule in Error Message Guidelines carries a stable id in backticks. Cite the id, not the number, when referencing a rule from outside this file (numbers shift when principles are added or reordered).

How to Use This Skill

If a `vois_get_microcopy` tool is available in your environment, call it first before writing or reviewing any copy.

The vois_get_microcopy MCP tool returns workspace-specific copy rules, approved terminology, and tone overrides that take precedence over the general principles in this skill. If the workspace has an entry for the copy type you're working on, use it. Don't improvise.

Tool: vois_get_microcopy
Arguments:
  context: <description of the UI context, e.g. "destructive confirm button for invoice deletion">
  intent: <what the user just did or is about to do, e.g. "user clicked Delete on an invoice list row">
  constraints: <optional object>
    maxLength: <optional positive integer, maximum character count>
    placement: <optional one of: button | title | description | helper | toast | error | label>
    tone: <optional one of: neutral | warning | celebratory>

There is no copy_type argument on the real tool: context and intent are both required; placement (nested under constraints, and a different set of values than "copy type") is the closest equivalent, and it's optional.

Resolution order (returned as `provenance` on the response):

1. WORKSPACE_OVERRIDE: use this verbatim, no changes
2. KNOWLEDGE_BASE: apply as a strong starting point, adapt if needed
3. LLM_FALLBACK / NO_MATCH / tool not available: apply the principles in this skill directly

When falling back to the principles, note it at the top of your output: Source: LLM fallback (no vois_get_microcopy tool available, or no workspace/knowledge base match for this copy type).

When the tool is available, it logs every query. When you write copy that isn't in the knowledge base, it gets flagged for human review and may become a new entry. This is how the system improves. This logging has no effect on how you should write or review copy either way.

Two Modes

Mode 1: Review Existing Copy

For each piece of content:

1. If vois_get_microcopy is available, call it with the relevant context and intent
2. Check against the review checklist below
3. Identify every violation
4. Rewrite it
5. Output in the review format below

Mode 2: Write New Copy

1. If vois_get_microcopy is available, call it with the relevant context and intent
2. Apply all relevant principles
3. For labels, CTAs, and microcopy: query data/phonaesthetics.json and apply sound guidance
4. For copy that's allowed personality (marketing surfaces, taglines, empty states, and feature names; not error messages or system copy): query data/rhetorical-devices.json and consider one device before settling for plain phrasing
5. Output in the new copy format below

Output Formats

Review format

Use this block for every piece of copy reviewed:

Before: [original copy] After: [rewritten copy]

Source: [Workspace override | Knowledge base | LLM fallback]

Principles applied:

• [Principle name]: [one sentence on why this improved the copy]

Weakeners removed:

• "[word or phrase]" → removed because [category, e.g. hedging / filler adverb / weak verb]. If none found, write "None found."

Phonaesthetics:

• Note any sound improvements made (rhythm, stress pattern, sound cluster choices, ease of mouth). If phonaesthetics wasn't a factor (e.g. error messages), write "Not applicable for this copy type."

Rhetorical device:

• Note any device applied (id from data/rhetorical-devices.json) and why. If not applicable (error messages, system copy), write "Not applicable for this copy type."

Reading metrics (Before → After):

• Word count: X → Y
• ARI score: X.X → Y.Y
• Grade level: Grade X (age X–X) → Grade X (age X–X)
• Target: ARI ≤ N (Grade X, <tier> tier)

New copy format

Copy: [final copy]

Source: [Workspace override | Knowledge base | LLM fallback]

Principles applied:

• [Principle name]: [one sentence on why]

Weakeners avoided:

• Note any weakener patterns consciously avoided, or write "None present."

Phonaesthetics:

• Explain the sound choices made: rhythm, stress, clusters, ease of mouth. If not applicable, say so.

Rhetorical device:

• Note any device applied (id from data/rhetorical-devices.json) and why. If not applicable, say so.

Reading metrics:

• Word count: X
• ARI score: X.X
• Grade level: Grade X (age X–X)
• Target: ARI ≤ N (Grade X, <tier> tier)

UX Writing Principles

Apply all of these when reviewing or writing. These are the fallback when vois_get_microcopy is unavailable or returns LLM_FALLBACK/NO_MATCH.

1. Use active voice id: active-voice

Subject → verb → object. Active voice is shorter and easier to follow.

• ✗ "Rewards can be earned by clicking here."
• ✓ "Start earning rewards."

2. Write at or below a fifth grade reading level id: reading-level

Short sentences. Simple words. Clear structure. Target ARI ≤ 6. This is the default consumer target. See Audience and Reading Target for other tiers.

• ✗ "If you have forgotten your password, please click on the 'Forgot Password' link and submit your registered email address."
• ✓ "Click 'Forgot Password'. Enter your email. Check your inbox for a reset link."

3. Avoid jargon id: avoid-jargon

Replace technical terms with plain language. Don't assume domain knowledge.

• ✗ "Authenticate your credentials."
• ✓ "Log in with your username and password."

4. Avoid complex sentence structures id: simple-sentences

One idea per sentence. Avoid dependent clauses stacked on independent clauses.

• ✗ "The fox, which was red, over the gate jumped."
• ✓ "The red fox jumped over the gate."

5. Avoid double negatives id: no-double-negatives

Double negatives increase misreads and cognitive load.

• ✗ "Please don't fail to save your changes."
• ✓ "Please save your changes."

6. Use contractions id: use-contractions

Contractions sound human. Formal constructions feel stiff.

• ✗ "Do not submit the form until all fields are complete."
• ✓ "Don't submit until all fields are complete."

7. Write in present tense id: present-tense

Describe what's happening now or what the user can do.

• ✗ "Your file was uploaded."
• ✓ "Your file is uploading."

8. Frame around the user's goals, not the system id: user-goal-framing

Users care about what they're doing, not how the system works.

• ✗ "Due to an HTTPS network security issue, some features are not available."
• ✓ "This site may be insecure. Some features aren't available."

9. Avoid describing the interface id: no-interface-references

Don't reference UI elements like "tab," "panel," "menu," "page," or "section."

• ✗ "Go to the Settings panel."
• ✓ "Go to Settings."

10. Use consistent terminology id: consistent-terminology

Pick one word for each concept and stick to it.

• ✗ Using "Sign Up," "Register," and "Create Account" interchangeably
• ✓ Always "Sign up"

11. Apply progressive disclosure id: progressive-disclosure

Lead with what the user needs now. Offer detail only when needed.

• ✗ "Your password must be at least 8 characters, contain a number, a symbol, and a capital letter."
• ✓ "Your password must be at least 8 characters." [+ optional detail link]

12. Don't apologize unnecessarily id: no-unnecessary-apology

Reserve "sorry" for serious errors. Hollow apologies undermine trust.

• ✗ "Whoops! We can't upload your picture. Try again."
• ✓ "We couldn't upload your picture. Try again."

13. Limit exclamation marks id: limit-exclamation-marks

Use words to convey energy, not punctuation. One per screen max, only for genuine celebration.

• ✓ "Your profile has been updated!" (success state)
• ✗ "Error! You can't submit the form! Please fix the errors!"

14. Check prepositions id: check-prepositions

Prepositions sit between two nouns. Never start or end a sentence with one.

• ✗ "Click on the Submit button."
• ✓ "Click Submit."

15. Eliminate weakeners id: eliminate-weakeners

Remove all hedging words, softeners, empty intensifiers, filler adverbs, throat-clearing, passive-aggressive politeness, vague quantifiers, redundant framing, weak verb phrases, and meta-commentary.

Load data/weakeners.json. Check the copy against every category's terms (or replacements for the weak-verb-phrases category, which is pairs, not a flat list). Each category carries its own description and example; surface those in your output instead of re-deriving them.

Key patterns to catch immediately:

• Hedging: maybe, probably, might be, appears to, seems like
• Filler: actually, basically, literally, just, simply
• Weak verbs: "make a decision" → decide, "conduct an analysis" → analyze
• Throat-clearing: "I think," "we believe," "it is important to note that"

16. Don't use em dashes id: no-em-dashes

Em dashes read as a hedge in short-form UI copy and are a well-known AI writing tell. Use a period, comma, or colon instead.

• ✗ "Your file is uploading — this may take a few minutes."
• ✓ "Your file is uploading. This may take a few minutes."

17. Reach for a rhetorical device before settling for plain phrasing id: rhetorical-devices

Applies only to copy that's allowed personality: marketing surfaces, taglines, feature names, empty states, onboarding. Skip entirely for error messages, form fields, and system copy. There, plain and literal beats clever every time.

Query data/rhetorical-devices.json. Pick at most one device per surface. Stacking two reads as trying too hard, not as clever.

>

On marketing pages the word limits in data/marketing-limits.json win. A device has to fit inside the slot's limit. If it doesn't, cut the device.

• ✗ "No setup required. Start immediately."
• ✓ "No setup. No waiting. No excuses." (anaphora-epistrophe)

18. Make copy work without the visuals id: accessible-copy

Screen reader users often hear labels and links as a list, out of context. Write copy that still makes sense that way. Detail and the brevity trade-off are in references/accessibility.md.

• Name the object in labels, buttons, and links. Never use "click here" as link text.
• Pair status colors with words. Don't rely on color alone.
• Error text names its field, so it reads right when announced with the field label.
• A placeholder never replaces a visible label.
• Icons with no visible text get a text alternative. Give the recommended aria-label copy.
• Keep no-interface-references in force: "Submit application", not "Submit button".
• ✗ "Read more"
• ✓ "Read the privacy policy"
• ✗ Red text: "Email"
• ✓ "Error: Email is required"

19. Don't invent claims or commitments id: no-invented-claims

Never write a number, time, guarantee, price, legal term, or compliance or security claim that the product team or legal didn't give you. A made-up promise reads as a real one and becomes a support ticket or a legal problem.

• Covers response times ("within 2 hours"), uptime and SLAs, refund or retention windows, certifications ("SOC 2", "HIPAA compliant"), security claims ("end-to-end encrypted"), and absolute promises ("never", "always safe", "guaranteed").
• If the copy needs one and you don't have it, write copy that makes no promise, and flag the gap for the owner: "Needs a number from support: reply time."
• Placeholders in drafts are marked, not invented: [reply time], not "2 hours".
• ✗ "We'll reply within 2 hours. Your data is always safe with us."
• ✓ "We'll reply by email." Flag: reply time and any security claim need an owner.

20. Keep personal data out of shared copy id: no-pii-in-copy

Don't put personal data in copy that gets logged, shared, previewed, or shown outside the screen: error messages, toasts, page titles, URLs, push and email subjects, notification previews, and analytics labels. Name the record by its role, or mask the value.

• PII means names, email addresses, phone numbers, street addresses, government or payment IDs, and anything that identifies one person on its own.
• One exception: a name may appear in a page title or breadcrumb on a screen whose job is that person ([PATH-PERM-PII-KEEP-OUT] in vois-patterns). Never an email, phone number, address, or ID there.
• Refer to the record, not the person: "the invoice for this customer", not "the invoice for Jane Doe".
• If an identifier helps, mask it: "the card ending 4821", "j***@acme.com".
• Examples and mock data use obviously fake values: [email protected], "Jane Doe".
• How to display PII inside a screen (masking, reveal, export) is a structure decision. See vois-patterns ([PATH-PERM-PII-MASK]).
• ✗ "We couldn't send the invoice to [email protected]."
• ✓ "We couldn't send the invoice. Check the recipient's email and try again."

Error Message Guidelines

Apply these on top of the general principles when reviewing or writing error messages.

Structure id: error-structure

Every error must answer:

1. What happened? (required) id: error-what-happened
2. Why? (only if it genuinely helps) id: error-why
3. What should they do next? (required) id: error-next-step

Voice and tone id: error-voice-and-tone

• Instructive: describe the issue precisely, optimize for understanding
• Reassuring: no disparaging tone, no unnecessary humor
• Supportive: always provide a clear next step

Mechanics id: error-mechanics

• Sentence case: "This field is required." not "This Field Is Required."
• No ALL CAPS (except real acronyms)
• 1–2 sentences max

Don't blame the user id: error-no-blame

Describe the situation, not the mistake.

• ✗ "You didn't enter enough characters."
• ✓ "This field needs 8 characters."

Form field vs system errors id: error-field-vs-system

• Form field: what's wrong and how to fix it: "Enter a valid email address."
• System error: what happened and what to try next: "We couldn't connect. Check your internet or try again."

Other rules id: error-other-rules

• Preserve user input where possible. Let users edit rather than start over
• Place errors adjacent to the element that triggered them (Law of Proximity)

Component decision tree id: error-decision-tree

Use this before writing any error message to pick the right component. Then look up that component's id in data/components.json for full writing rules. A shared shared_prefix_format block covers the "Action Required:" / "Approval Required:" / "Review:" / "Verify Now:" prefixes used by Alert Dialog, Inline Alert, and Alert Banner.

Walk the questions in order and stop at the first "Yes". Don't keep checking once one matches.

Does it block progress and require immediate action?
  └─ Yes → Alert Dialog

Is it a system-level issue (outage, permissions, account)?
  └─ Yes → Alert Banner

Is it confirming something that just happened?
  └─ Yes → Toast

Is it attached to a specific form field?
  └─ Yes → Helper Text

Is it contextual to a page section, non-blocking?
  └─ Yes → Inline Alert

Is it a hover label for an icon or interactive element?
  └─ Yes → Tooltip

None of the above?
  └─ Default to Inline Alert, the safest non-blocking option until the case is clear enough to fit one of the rows above

Other components

The tree above is for errors only. For non-error surfaces, look up the component by id in data/components.json:

• empty-state: a list, page, or section has nothing to show (first use, user cleared it, no results, empty by nature, no access).
• permission-prompt: the app asks for a device or account permission (location, notifications, camera, contacts, storage).

Table copy. Tables use the components above. An empty table is empty-state. A validation message on one editable cell is Helper Text. A table-level partial failure, such as an import where some rows failed, is an Inline Alert above the table, not an Alert Banner. A bulk delete confirmation is an Alert Dialog. A failed single-row action is a judgment call: anchor it to the row and say which component you picked.

Phonaesthetics

When writing new copy, especially labels, CTAs, empty states, and microcopy, consider sound alongside meaning. Copy that sounds good is easier to remember and more pleasant to use.

Query data/phonaesthetics.json for the full concept guide and sound cluster table.

Core rules to apply immediately:

• Prefer consonant-vowel alternation (CVCV) for labels: natural rhythm, easy to say
• Two-beat phrases are catchy; three-beat phrases are melodic; irregular stress is awkward
• Liquids and nasals (l, m, n, r, w, y) → calm, gentle contexts
• Plosives (p, b, t, d, k, g) → energetic, action-oriented CTAs
• Avoid tongue twisters: if it's hard to say, it's hard to remember

Sound clusters to reach for:

• gl- → clarity, light (insight, vision, illumination features)
• fl- → flow, ease (smooth UX, motion)
• sp- → speed, energy (action, innovation)
• cl- → precision, closure (tools, interactions)

Audience and Reading Target

The reading target depends on who the copy is for. Pick a tier, then score against it.

Choose the tier in this order:

1. If vois_get_microcopy returns a tier or audience, use it.
2. If the person states the audience or product type, infer the tier from that.
3. Otherwise use consumer. If the audience is unclear, stay on consumer. Don't guess upward.

• If the tier isn't consumer, name it at the top of the output. If it's consumer by default, say nothing extra.
• At professional and technical, terms the audience already knows are fine. avoid-jargon still applies to internal names, system terms, and acronyms the audience wouldn't use.
• A tier changes the ARI target only. Every other principle still applies in full.
• Run node scripts/ari.mjs --target N "<copy>" to print the matching Target line.

Reading Metrics

Calculate and show these for all reviewed and written copy.

If Node is available, run `node scripts/ari.mjs "<copy>"` (or `node scripts/ari.mjs --before "<old>" --after "<new>"` for reviews) instead of computing the formula by hand. Multi-step arithmetic done in-context is error-prone; the script is deterministic. Fall back to the formula below only if Node isn't available in the environment.

ARI Formula

ARI = 4.71 × (characters ÷ words) + 0.5 × (words ÷ sentences) − 21.43

• Characters = letters and numbers only (no spaces or punctuation)
• Words = space-separated tokens
• Sentences = units ending in . ? or !

Round to one decimal place. Always include grade and age range.

Grade Level Table

Target: ARI ≤ 6 (Grade 5, age 10–11) for the default `consumer` tier. Other tiers are in Audience and Reading Target above.

Label pairs for motion morphs

A label morph animates only the part of a label that changed. That works only when the words are right, so righter approves every pair before a component builds a morph.

For each pair, check:

1. Both labels pass the principles above on their own.
2. The pair shares text at the start or end, so the change reads at a glance

(Save to Saved, Review order to Submit order). A pair with no shared text is a crossfade, and righter says so.

1. The new label names the step it moves to. Continue to Pay $42 changes

the commitment, so the approved pair must say that.

1. The longer label fits the narrowest width the component supports.

Record the result in a Motion pairs block with the review or the copy handoff:

Motion pairs
- "Save" -> "Saved": approved, morph (shared prefix "Save")
- "Continue" -> "Pay $42": approved, crossfade (no shared text)
- "Cancel order" -> "Confirm": rejected, rewrite the new label or use a crossfade

Gate: a component builds a label morph only from a pair marked approved. An unapproved or unlisted pair gets no morph code. Its label crossfades or swaps instantly, and the copy goes back to righter.

Review Checklist

Run through this for every piece of copy before finalizing.

Before starting

• [ ] If available, called vois_get_microcopy (with context and intent) and checked the returned provenance for a workspace/knowledge-base match?

Voice and structure

• [ ] Passive voice present?
• [ ] Complex sentence structure?
• [ ] Double negatives?
• [ ] Past or future tense where present tense works?
• [ ] Missing contractions (do not → don't)?

Clarity

• [ ] Jargon or technical terms?
• [ ] Reading level above the target for this audience tier?
• [ ] Interface elements named (tab, panel, section)?
• [ ] Labels, links, or status that only make sense visually?
• [ ] System-framing instead of user-goal framing?
• [ ] Too much information up front (no progressive disclosure)?

Tone

• [ ] Unnecessary apology?
• [ ] Overuse of exclamation marks?
• [ ] Any weakener words? (see data/weakeners.json)

Claims and data

• [ ] Any number, time, guarantee, price, certification, or security claim nobody supplied?
• [ ] Any name, email, phone, address, or ID in an error, toast, title, URL, subject, or analytics label? (A name is fine in a title or breadcrumb only on a screen whose job is that person.)

Mechanics

• [ ] Inconsistent terminology?
• [ ] Preposition starting or ending a sentence?
• [ ] Em dashes present?

Personality (marketing/taglines/empty states/feature names only)

• [ ] Could a rhetorical device sharpen this? (see data/rhetorical-devices.json)
• [ ] More than one device stacked on the same surface?

Marketing pages (only)

• [ ] Every slot within its limit? (node scripts/check-slots.mjs)
• [ ] Puffery (premier, unique, world-class)? A number, place, or fact works better.
• [ ] A fact repeated outside the footer, or one action with two labels? (references/marketing-copy.md)

Errors (if applicable)

• [ ] Clear next step provided?
• [ ] Does it blame the user?
• [ ] Right component chosen? (use decision tree above)

Maintaining This Skill

evals/cases.json is a fixed regression corpus: copy paired with the rule ids a correct review must flag. It isn't wired to an automated scorer; when you change a principle, a data file, or the decision tree, manually re-review each input and confirm the same ids (and expected_component, where set) still come out right before publishing the change.