| name | unslop-copywriting |
| description | Bans AI-slop public UI copy and rewrites website/app text into plain language for humans. Removes empty marketing verbs, robotic microcopy, vague CTAs, apology-padded errors, and especially leaked internals — agent instructions, implementation criteria, schema/API names, TODOs, stack traces, and prompt residue rendered as user-facing strings. Use when the user says "unslop-copywriting", "unslop copy", "AI copy", "robotic copy", "plain language UI", or asks to fix microcopy, marketing copy, errors, empty states, or text that reads like it was written for machines. |
Unslop Copywriting
AI models write public UI text in two failure modes: marketing slop (empty empowerment verbs, interchangeable SaaS claims) and boundary failure (instructions, criteria, schema, and process notes rendered as if they were product copy). This skill bans both and rewrites until every user-visible string reads as plain language for a human completing a task.
Core Mandate
- Public UI is for humans. Never ship copy that sounds like a system prompt, ticket, schema, stack trace, or implementation note.
- Ban first, then rewrite. Do not polish slop in place. Strip banned phrases and leaked internals, then write specific plain language.
- Prefer delete. If a string adds no task value, remove it. Models always produce text; humans often should not.
- Outcome over fluff. Name the concrete action, destination, or result. Vague verbs and ghost CTAs are failures.
- Pivot rule. If a surface hits 3+ banned patterns, rewrite the whole message set for that surface — do not synonym-swap.
Hard Bans (never ship)
| Cluster | Banned |
|---|
| Leaked internals | Agent/dev instructions as headings or body; “criteria” / permission / fallback notes as visible copy; API paths, schema/JSON keys, feature flags, env names; TODO / FIXME / lorem ipsum / placeholder honesty; stack traces, raw error.message, HTTP codes as headlines; “As an AI…”, session narration, prompt residue |
| Marketing voice | unlock, empower, elevate, supercharge, harness, unleash, revolutionize, transform (without concrete before→after); effortlessly, seamlessly, streamline; empty adjectives (powerful, robust, cutting-edge, next-generation, AI-powered as filler); Welcome to our platform, The future of X, Built for modern teams |
| Rhetoric | It's not X — it's Y / Not just X, but Y; In today's …; Let's dive in; Furthermore / Moreover / It's worth noting; em-dash abuse in short UI strings; rule-of-three adjective stacks (Fast. Secure. Reliable.) |
| Controls / CTAs | Lone Submit, OK, Click here, Learn more, Get started, Yes/No when an outcome verb exists; emoji or ! on buttons |
| Errors / empty | Something went wrong / An error occurred alone; Invalid input; blame (illegal, you must, ALL CAPS); apology padding on validation; fake “we've been notified”; empty states with no next step; No data available database voice |
| Tone | Over-formal committee voice on routine UI; forced cheer (Awesome!!) on high-stakes flows; humor in errors; sensory-only directions (click the green button) |
Full catalog: references/banned-patterns.md.
Workflow
1. Scope
- Full unslop: all first-party user-visible strings (pages, components, toasts, emails in-app, i18n/locale files,
aria-label, empty/error states, marketing heroes).
- Targeted unslop: named routes, components, locales, or copy categories only.
State the scope, then proceed. Do not rewrite legal/compliance text unless asked; flag it if it fails plain-language tests.
2. Audit
Search JSX/HTML/i18n for banned phrases, leak markers, and vague CTAs. Record file, string, pattern id, and severity.
Critical severity (must fix): any leaked internal (I*), stack/exception in UI, prompt residue.
Use references/review-checklist.md.
3. Lock a voice contract
Before rewriting, write a short contract:
- Audience and reading level (default: plain language, ~6th–8th grade for general UI)
- Product vocabulary glossary (one word per concept: Sign in vs Log in — pick one)
- Voice (plain, direct, specific) and tone shifts by stakes (calmer on billing/errors; warmer on success)
- What must never appear (paste hard bans that matter for this product)
- When silence is better than help text
If the contract would fit any CRM/mattress/tax tool unchanged, revise it.
Craft rules: references/craft-antidotes.md.
4. Rework
For each banned hit:
- Delete leaked internals and filler — do not paraphrase instructions into “nicer” instructions.
- Rewrite remaining copy in plain language: specific verb, user words, one idea per sentence.
- Buttons: verb (+ noun) naming the outcome; sentence case; no punctuation/emoji.
- Errors: what happened → why (if useful) → what to do next. Map exceptions via a user-message helper; never render raw errors.
- Empty states: why empty → one primary next action. Distinguish first-run vs failure-as-empty.
- Marketing: replace every vague adjective with a number, every adverb with a comparison, every power verb with a concrete action — or cut the claim.
- Preserve meaning and behavior; change copy only. Ask before altering legal/brand-mandated strings.
5. Verify
Re-run the checklist. Pass the human test: read aloud — would a non-engineer know what to do next with no internal knowledge? Pass the any-product test: if the same sentence could sit on an unrelated product unchanged, rewrite.
Completion Report
- Scope audited
- Leaks removed (critical)
- Banned patterns rewritten (by cluster)
- Voice contract locked
- Remaining exceptions (brief-justified only)
- Checklist status
Reference Routing
