Better WritingSAFE

A collection of agent skills that help you build great interfaces.

The instruction file, verbatim from the audited commit — this is the text the model reads, and the surface the audit's instruction layer examines. Quoted here so you can judge it without cloning anything.

---
name: better-writing
description: Writes and reviews interface copy, from labels and errors to empty states and confirmations, so it matches the product's voice and tells people what to do next.
---

# Writing

This skill writes and reviews interface copy such as labels, errors, empty states and confirmations, along with the terms that run through them. It matches the product's existing voice and flags copy that misleads people or leaves them without a way forward.

A deliberate brand voice is not a defect. Raise a departure from plain language only when it creates inconsistency, ambiguity, translation risk or a tone the stakes don't support. Rewording that merely suits your taste is not a finding.

How copy renders, including `text-transform`, truncation and smart punctuation, belongs to `better-typography`. Error markup, announcements and the attributes that carry accessible names belong to `better-accessibility`. Room for translated strings belongs to `better-layout`.

## Inventory the existing strings first

Before writing or reviewing, find where the copy lives and read the copy around the change:

1. Search for the translation call (`t(`, `i18n.`, `<FormattedMessage`, `$t(`) and the locale files it reads, such as `locales/**/*.json`, `messages/*.json`, `*.po` and `*.strings`. Copy may also come from a CMS or the API.
2. List the noun used for each object and the verb used for each action, as in "project" or "workspace" and "Delete" or "Remove".
3. Note the case used per element type and any voice or content style guide.

New copy uses the terms on that list. A synonym is a finding only where the same thing is named two ways.

## One voice, one vocabulary

The product has one voice and its existing copy establishes it. A local edit does not get to invent a new one. If it's "Archive" in the menu, it isn't "Move to storage" in the toast.

A multi-step flow uses one vocabulary throughout: "Get started" to enter, either "Continue" or "Next" to advance, "Done" to finish. Tone flexes with the stakes:

| Context | Tone |
| --- | --- |
| Success, onboarding, empty states | Warm, can be light |
| Routine actions, settings | Neutral, minimal |
| Errors, destructive confirmations | Calm, plain, zero playfulness |
| Data loss, security | Serious, explicit |

## Address the reader directly

In instructional copy, write "you", not "the user". In errors, "we" reads as deflection, so prefer "Unable to load content. Check your connection and try again." An established first-person voice can stay in low-stakes copy where it still reads clearly.

Use possessives sparingly: "Favorites" beats "Your favorites". Never mix perspectives in one flow, such as "My account" beside "Your settings".

## Plain words over clever ones

Choose words a tired reader gets on the first pass, and delete every word that does no work. No idioms, no colloquialisms and no humor that won't translate.

Skip unnecessary gender: "Subscribers can post recipes", not "each subscriber can post his or her recipes". Match the input device: "tap" on touch, "click" with a pointer, "select" when both are possible.

## Build strings whole

Never assemble a sentence from fragments around a variable (`"You have " + n + " new messages"`), because word order changes per language. Write one message with placeholders, and use the locale's plural rules rather than `n === 1 ? "" : "s"`. Many languages have more than two plural forms. Format numbers, dates and lists through `Intl`. Recipes are in [patterns.md](patterns.md#plurals-and-placeholders).

## Verb-first buttons

A button label starts with a verb naming the action: "Send", "Save draft", "Delete project". Never "OK!" or "Let's go!".

An icon-only button's accessible name follows the same rule. It names the action, "Delete project", never the glyph, "Trash icon". The attribute that carries it is `better-accessibility`'s.

## Links describe their destination

Link text makes sense out of context, because screen-reader users navigate by a list of the page's links. Write "Read the billing docs", never "Click here".

A bare "Learn more" breaks down as soon as two appear on one page. Suffix each one: "Learn more about exports".

## One capitalization policy

Pick title case or sentence case per element type, then apply it to every instance of that type. Sentence case is the default where the project has no policy. "Save Changes" beside "Discard changes" reads as sloppiness.

## Settings describe the ON state

Label a toggle for what happens when it is on. "Send read receipts" lets users infer the off state; "Don't send read receipts" turns the toggle into a double negative.

Link straight to a referenced setting rather than describing the path to it: a "Notification settings" link, not "Go to Settings > Notifications > Email".

## Errors say how to fix, next to where it broke

An error is an instruction, and it belongs beside the field that failed:

| Bad | Good |
| --- | --- |
| That password is too short | Choose a password with at least 8 characters |
| Invalid date | Enter a date as DD/MM/YYYY |
| Oops! Something went wrong. | Unable to save. Check your connection and try again. |

No blame, no "oops" and no exclamation marks. Phrase hints positively, as in "Use at least 8 characters" rather than "Don't use fewer than 8". Show a known requirement as helper text before input, not only in the error. When the same error keeps firing, redesign the interaction so it cannot happen.

## Undo beats confirmation

Prefer undo when the action can be reversed and people perform it often, such as archiving, moving or deleting into a trash. Act at once and offer undo in the result: "Project archived. Undo".

Confirm before acting when the action cannot be reversed, affects other people or destroys many items at once. The confirmation repeats the consequence, so the dialog is answerable without reading the body:

- The title names the action and the object, "Delete 'Q3 report'?", never "Are you sure?".

SAFEgrade B · trust 89/100 Nothing in the source contradicts what it says it does. Grade A is reserved for packages that have also passed the behavioural sandbox.

What the source does

Filesystem

none-observed

Network

none-observed

Shell

none-observed

Dependencies

pinned

Secrets in source

none-found

Findings (0)

No findings outside the package's declared scope.

Gates applied: no_behavioural_pass.

[![ModelsAtlas audit: Better Writing](https://modelsatlas.com/skills/jakubkrehel/better-writing/badge.svg)](https://modelsatlas.com/skills/jakubkrehel/better-writing/)

Not your listing, or something out of date? Claim or update your listing.