copywriting

name	copywriting
allowed-tools	Read Write Edit Grep Glob WebFetch Bash(python3:*)
description	Authors and judges copy.yaml, the content payload a design consumes: write fresh brand or product copy from intent, or extract and structure existing content from a URL, brief, codebase, or screenshot — content only, composable under any design — then critique or audit it before it ships. Use when writing headlines, value propositions, landing-page or CTA copy; extracting, structuring, tightening, or revoicing content; capturing copy from a screenshot; syncing copy.yaml from a drifted implementation; critiquing copy or checking whether it reads as AI slop; scoring copy on voice and proof; or auditing copy for readability, claims, and microcopy before handoff. Not for visual identity or design tokens, page layout or screen flow, or standalone social bios.

Copywriting

Owns copy.yaml — the structured content payload a design consumes. Content is orthogonal to design: the same copy.yaml must drop into any visual identity, so this skill carries words only, never design decisions. It both authors copy and judges it; the judging modes are non-mutating — they report, the authoring modes apply.

Quick start

Author — produce or change copy:

• write — author fresh copy from intent (headlines, body, CTAs). → write.md
• extract — structure existing content from a source (URL, brief, codebase, screenshot), preserving tone. → extract.md
• refresh — tighten existing copy in the same voice (editing passes). → refresh.md
• revoice — rewrite existing copy in a new voice, keeping the message. → revoice.md
• reconcile — sync copy.yaml from a drifted implementation. → reconcile.md

Judge — a non-mutating verdict on existing copy:

• critique — quality and slop verdict on a draft; scores the seven sweeps, loops back to refresh. → critique.md
• audit — ship-readiness defect report on copy.yaml, P0–P3, before handoff. → audit.md

Discovery

discovery.md runs before every operation — never skipped, never invoked directly. It checks existing context (copy.yaml, source, upstream intent), classifies the request (author vs judge, then greenfield vs brownfield for authoring), and routes to the matching operation. See discovery.md.

Artifact

Produces and owns docs/design/copy.yaml — a context-named content tree (surfaces → parts — headline, body, cta, labels, states, images — named by context), mirroring the source. Before saving, self-check: the tree is well-formed and carries no design decisions — no colors, fonts, or layout, content only. The content stays swappable: any copy.yaml must work independent of visual styling.

Register and surface

Two axes, set before writing:

• register — the posture: brand (the words are the product) or product (the words serve the task). Sets the voice. Read the matching brand.md / product.md first.
• surface — the granular type the copy serves, named by context (landing, dashboard, form, empty-state…). A surface sits under a register; the content tree is named by context, never forced into a fixed list. Storefronts straddle — catalog copy is brand, checkout / account copy is product.

References

Loaded on demand by the workflows:

• references/brand.md / references/product.md — register (brand vs product) posture; read the matching one first
• references/copy-frameworks.md — headline formulas, content-part types, page shapes, CTA patterns
• references/voice.md — how register sets the voice, voice axes, proof hierarchy
• references/editing-sweeps.md — Seven Sweeps, quick-pass checks, plain-English
• references/ux-writing.md — clarity craft: the assess→plan→improve→verify method, clarity principles, microcopy (errors, labels, states), a11y/i18n/terminology
• references/anti-patterns.md — copy slop catalog: dead words, dead structures, AI tells, proof failures; composed by both author and judge modes
• references/scoring.md — shared severity, score bands, and report template for critique and audit
• scripts/slop_scan.py — run for the deterministic slop tally (dead words, em-dash density, openers) that critique and audit consume
• scripts/validate_copy.py — run after any write/patch for the deterministic well-formedness and design-leakage scan the authoring self-checks rely on

Non-mutating judgment

critique and audit read and report — they never patch copy.yaml. To apply a verdict, run the matching authoring operation: a weak critique axis loops to refresh, an off-register voice to revoice, a missing part to write — each confirmed before write. The judging modes produce the verdict; the authoring modes own the change.

Guidelines

• When writing, be specific and keep proof outward — a number, name, or example beats an adjective.
• Preserve the source's tone when extracting — structure content, do not rewrite it.
• Keep copy.yaml content-only; never embed visual decisions.
• When judging, name the register first and lead with the slop verdict; let the score support it, not replace it.
• Scope output to what was captured — a region produces a region, not a full-surface tree.
• Name surfaces and parts by context; mirror the source's own structure.