Web Copy

Improves or drafts website copy that reads like it came from a senior copywriter in a digital agency: clear, persuasive, friendly, SEO-sound, in European English.

Two modes, one rubric:

• Review — the user pastes copy (or a URL or a file path). Score against the rubric, mark up issues, return a clean rewrite.
• Draft — the user gives a brief. Pick a framework, draft the page, then self-check against the same rubric before returning.

Default tone is professional-friendly (Mailchimp / Shopify Polaris territory). Override with a tone: line in the brief if the user wants formal or playful.

Language default is en-GB. Never emit -ize spellings, color, favorite, gotten, awesome, reach out, single-word Americanisms, or the Oxford comma (unless it removes real ambiguity).

Process

1. Decide the mode

If the user pasted or pointed at existing copy → review. If the user described a page they want written → draft. If ambiguous, ask once.

For a URL, fetch it with the built-in WebFetch tool and pull the visible copy before scoring. For a file path, Read it.

2. Load the rubric

Read references/rubric.md before scoring or writing anything. Every finding, every rewrite decision, has to map to a rubric line. This is the anchor that keeps output consistent between runs.

Load these on demand — do not load them upfront:

• references/european-english.md — every run (en-GB is the default).
• references/seo-checklist.md — for any page that will be indexed.
• references/frameworks.md — before drafting or when a rewrite needs a structural fix, not a line-level fix.
• references/microcopy.md — for buttons, forms, errors, empty states, 404s.
• references/page-patterns.md — when the brief is "hero", "pricing", "about", "features", "FAQ", "testimonials", or "footer".
• references/authors.md — when justifying a specific move (cite the author the rule comes from, e.g. Schwartz on awareness, Sugarman on the slippery slide, Ogilvy on headlines). Adds credibility, keeps the skill from inventing rules.
• references/sources.md — when the user asks why a rule exists or wants further reading.
• references/llm-fingerprints.md — every draft and every review, not optional. Enforcement backbone for category 9 (voice authenticity).

3. Review mode — the output shape

Return in this order, in one message:

1. Verdict — one sentence. What's the biggest problem, or what's working.
2. Findings table — Markdown table with columns: Location (which sentence / heading / CTA), Issue (rubric category), Severity (high / medium / low), Suggested rewrite.
3. Clean rewrite — the full copy, rewritten, ready to paste. No commentary inside the rewrite.
4. SEO block — proposed <title> (≤ 60 chars), meta description (≤ 155 chars), suggested H1, target keyword, and any internal-link opportunities you spotted.
5. What I did not change and why — anything the user might expect to have been touched but wasn't (e.g. a legally required disclaimer).

Use the layout in assets/before-after.md as the template.

4. Draft mode — the output shape

1. Chosen framework — one line. "Using PAS because the audience is problem-aware but not solution-aware (Schwartz stage 2)." If no framework fits, say so and write plainly.
2. The draft — the full page copy, with headings marked (H1/H2/H3), body paragraphs, CTAs, and any microcopy the page needs.
3. SEO block — same fields as review mode.
4. Self-check — a two-line pass confirming en-GB, one H1, tone, active-voice ratio, no unverifiable superlatives, and no LLM fingerprints (see references/llm-fingerprints.md).

If the brief is missing something critical (audience, primary action, offer, brand voice), ask once, list the missing items together, then proceed with sensible defaults if the user says "your call".

Use the layout in assets/brief-template.md for what a good brief contains.

5. Always-on rules

• One H1 per page. More than one is an SEO bug, not a style choice.
• Benefits before features. Lead every section with what the reader gets, not what the product does.
• Specific numbers over vague claims. "Cut invoice time from 40 minutes to 4" beats "save time".
• No unverifiable superlatives (best, world-class, leading) unless the user supplies proof. If they do, cite it inline.
• No stock openers. No "In today's fast-paced world", no "We are excited to announce", no "Welcome to our website".
• Active voice ≥ 80%. Passive is fine when the agent is genuinely unknown or unimportant.
• Contractions on for professional-friendly. Off for formal.
• Exclamation marks off unless the user's brand voice explicitly wants them.
• CTAs are verb-led (Get my quote, Start free, See the demo). Not Submit, Click here, or Learn more.
• One primary CTA per section. Secondary CTAs may exist but must not compete for weight.
• Voice authenticity check. Before returning any draft or rewrite, scan against every pattern in references/llm-fingerprints.md. Every fingerprint is a fail — fix in place, don't ship.

Edge cases

• User pasted non-copy (code, JSON, a brief instead of the copy) — ask which mode they want.
• User asked for US English — honour it, but flag once that the skill defaults to en-GB.
• User asked for another language — say the skill is en-GB only; offer to do the strategic review (rubric, structure, SEO) and leave the language version to them.
• Legal, medical, financial claims — do not soften or invent them. Flag any claim that reads as regulated and ask the user to confirm the source.
• User's brand voice already exists (style guide, tone doc) — read it, defer to it wherever it conflicts with the skill's defaults, and say so in the verdict line.
• Very short copy (a single CTA, a headline) — skip the findings table; return the rewrite plus one line of rationale.

Checks before finishing

• Every finding in the table maps to a rubric line in references/rubric.md.
• The clean rewrite is self-contained — a reader could paste it into the page without editing.
• SEO block is present for any page-length rewrite (title, meta, H1, keyword).
• en-GB check has been run: no -ize, no color, no favorite, no Americanisms.
• One H1 in the rewrite. Not zero, not two.
• Every superlative is backed or removed.
• Voice-authenticity scan run: em-dashes ≤ 1 per 100 words, no double-"I" clauses without a third repetition, no AI-slop vocabulary (supercharged, AI-powered, harnessing AI, etc.), no "As an X" openers, no "In conclusion" closers, no X-not-Y aphoristic paragraph closers. Full list in references/llm-fingerprints.md.