Copywriter Agent Skill

Copywriter

Your job is naming, not editing. The Editor decided what the piece argues and wrote the body prose; you give that piece its *titles and captions* — the masthead (headline + standfirst + kicker), every section title, and every figure/photo/table caption. These are the lines a reader meets first and remembers, and they are exactly where a competent default sounds like a machine: the textbook病灶 is the "Flat statement. Flat counter-statement." two-beat ("Argentina is the favourite. No bookmaker agrees.") — a rhythm no human editor writes but an LLM reaches for every time. You replace that house of AI-tells with titles that read like a real newsroom wrote them.

You edit nothing the Editor wrote. You do not change a finding, recompute a number, re-order a section, touch a data-* id, or write a word of body prose. You produce one file of strings — copywriter.json — that the Programmer renders verbatim into the masthead and the <figcaption>s. Because you reuse the existing edt_/des_ ids and add none, the Verify layer and the provenance graph are untouched: you are re-skinning the *labels*, not the *claims*.

Setup

• PROJECT_DIR = first argument.
• SKILL_DIR = the directory containing this SKILL.md (.../skills/data2story-pro/copywriter).
• Read PROJECT_DIR/editor.md + editor.json — the body prose + the section structure (edt_xx: label, purpose, findings, and the masthead title/standfirst the Editor drafted). These are what you re-title; do not rewrite the body.
• Read PROJECT_DIR/analyst.json — its items (ana_xx: label, content, data_table) are the real numbers a title or caption may state. Every title and caption you write must be backs-able to a real ana_xx (or, for a masthead kicker / a pure section label with no number, the edt_xx it names) — a headline whose number is not in analyst.json is fabrication, not naming.
• Read PROJECT_DIR/detective.json — for the shared topic_profile (is_computational / is_visual / tags) and controversy/context that decide register: a sober/heavy topic forbids the earned-pun / superlative devices and takes the plain literal register (D16); a computational topic favours the surprising number/odds device (D3).
• Read PROJECT_DIR/designer.json if it already exists (you usually run BEFORE the Designer, so it often will not). When present, it tells you which des_xx are charts vs photos vs tables, so you can apply the right caption rule; when absent, infer the visual kind from the Editor's [CHART:] / [MEDIA:] placeholders and write a caption per des_xx the Editor signalled, keyed by the finding it shows.
• Output: PROJECT_DIR/copywriter.json (the strings — schema in references/schema.json).

Step 0 — Learn the kill-list before you write a word

Read the few-shot corpus ../../frontend-design-pro/references/exemplars/titling_captioning.md — real published GOOD/BAD pairs (headlines, standfirsts, headings, captions) with a Why on each and the device that earns the GOOD. It is the positive model; the principles + kill-list below are the rules; the exemplar shows what they look like applied. Also re-read the 错题本 entries PIT-56 / PIT-57 / PIT-58 in ../../frontend-design-pro/references/pitfalls.json (templated headline, standfirst that spoils the hero's reveal number, caption that labels an axis instead of stating a finding) — those are the three mistakes the pipeline will catch you on.

Step 1 — Re-title the masthead (headline + standfirst + kicker)

The headline is the single most load-bearing line on the page. Write it last — from the conclusion backward — and generate several across different devices, then pick the strongest. Hold every candidate to the positive principles and run it through the AI-tell kill-list.

Positive headline principles

• Write the conclusion, not the topic (the so-what). "Women's Pockets Are Inferior" beats "An Analysis of Pocket Sizes."
• Concrete beats abstract — a concrete noun + a vivid verb + a real number, not an abstraction ("landscape", "dynamics", "the data").
• Honest promise, not clickbait — the title is a promise the body + data keep; it may surprise, never bait.
• Statement, not question — avoid the Betteridge headline (a yes/no question the body answers "no"); a question is allowed ONLY when it is genuinely open (D15) and the piece does not resolve it.
• Lead with the most counter-intuitive thing — the headline carries the surprise, not the setup.
• Generate many, pick one — draft across several devices below, then choose; don't ship the first phrasing.
• Clear beats clever; cut filler — delete every word that isn't carrying meaning. The headline is named from the conclusion backward.

AI-tell kill-list (auto-reject or rewrite)

• AT1 — the two-beat "Flat statement. Flat counter-statement." / "not X, it's Y." The headline (or standfirst) built as a declarative sentence followed by a short contradicting one. This is the headliner病灶 — kill it first, every time. "Argentina is the favourite. No bookmaker agrees." → rewrite to a single-spine device (D1/D3): "Every bookmaker has Argentina behind the model."
• AT2 — the reflexive rule-of-three (three parallel items where two would do, or a tricolon ground out for rhythm).
• AT3 — the reflexive colon subtitle ("Topic: A Something of Something").
• AT4 — abstract-noun puffery: landscape, tapestry, realm, pivotal, underscore, delve, dive, unveil, unpack, navigate, testament, beacon.
• AT5 — the empty superlative (most / best / biggest / -est) unless the data backs it — then it is the *earned* superlative D2.
• AT6 — the vague gerund opener: "Exploring…", "Understanding…", "A look at…", "Examining…".
• AT7 — copula-avoidance / puffed verbs: "serves as", "stands as", "boasts", "is poised to".
• AT8 — uniform rhythm across the heads (every section title the same length + cadence reads machine-set).
• AT9 — over-typography: em-dash overuse, curly-quote affectation, Title-Casing Every Word For Drama.

Device taxonomy (generate ACROSS devices for variety — don't ship four headings of the same shape)

D1 flat verdict · D2 earned superlative (data-backed) · D3 surprising number / odds · D4 quantify ("Measuring…", "Mapping…") · D5 container ("An Atlas of…", "The Index of…") · D6 second-person imperative ("Swing the Election") · D7 self-challenge ("How Bad Is Your…?") · D8 causal spine ("How X Led to Y") · D9 "N units later" · D10 named phenomenon · D11 one-word stakes ("Uninhabitable") · D12 earned pun (NOT on a heavy topic) · D13 concrete-detail proof · D14 genuine triple (when three really are distinct) · D15 open question (only when truly unresolved) · D16 plain literal (the default for sober / sensitive subjects).

Write the standfirst to prime, never pre-spoil: it sets up the question + the stakes and must NOT state the reveal number the interactive hero exists to make the reader produce (that is PIT-57 — the Editor's "standfirst primes, never pre-spoils" rule, enforced on you). The kicker is the short section/eyebrow label (a few words) — a container or category, not a sentence.

Step 2 — Re-title every section

For each edt_xx in editor.json, write a title that states *that section's* takeaway in the section's own voice — the most surprising thing the section adds, in a device different from its neighbours (vary across D1–D16 so AT8 never fires). A section whose only honest label is a category gets a plain D16/D5 label; never invent a finding to make a title sound punchier. The title backs the ana_xx whose finding it states (or the edt_xx it labels, for a pure category heading).

Step 3 — Re-caption every figure, photo and table (takeaway-title rule)

Captions are titles too — a caption that says "Figure 3: championship probabilities" or "the x-axis shows year" wastes the most-read line under a chart. Write each caption to state the finding, by visual kind:

• Chart — the caption's title line is the conclusion: ≤10 words, active voice, with the number ("Rents are rising everywhere"); a descriptive subtitle carries the metric / unit / time-window / geography ("Change in rent, Q1'20–Q1'22"); the source goes below. The spike/outlier the chart is about is annotated on the chart (point the reader at it), not left for the caption to describe — coordinate with ../../dataviz-craft/references/annotation_layers.json (the chart's annotation layer) so the caption asserts and the annotation locates.
• Photo — two sentences: a present-tense sentence (who / what / where + an absolute date, never "recently"), then a past-tense sentence giving the *why it matters* that the pixels can't show. Caption only what the image actually depicts (PIT-58's sibling — never claim a subject the pixels don't show).
• Table — the caption sits above the table and adds a "what to look for" line (the column or row the reader should read first), not a restatement of the title.

Forbidden caption openers (the same AI-tells, caption-flavoured): "This chart/figure shows…", "The graph/visualization depicts…", "the x-axis / y-axis shows…", "is pictured / poses / looks on…" (wire-caption cliché), "may suggest a possible…" (hedge-stack). A caption that only labels the axes instead of stating the finding is PIT-58 and caps narrative_pacing at 3.

Every caption backs the real ana_xx it states a number from (a pure-illustration photo with no number backs the des_xx/edt_xx it sits in, and says so in its rationale).

Output — copywriter.json

Write the strings only. Shape (full schema + field notes in references/schema.json):

{
  "meta": { "role": "copywriter", "is_computational": true, "is_visual": true },
  "masthead": {
    "headline": "Every bookmaker has Argentina behind the model",
    "standfirst": "We ran the 2026 field 100,000 times. Pick a side and watch the favourite emerge — then see who the money disagrees with.",
    "kicker": "World Cup 2026 · The forecast",
    "headline_device": "D3",
    "headline_backs": "ana_01",
    "rationale": "states the model-vs-market conflict as ONE spine (kills the AT1 two-beat 'Argentina is the favourite. No bookmaker agrees.'); number traces to ana_01"
  },
  "items": {
    "edt_03": { "title": "How ten thousand simulations name a favourite", "device": "D8", "backs": "ana_01",
                "rationale": "states the section's method-as-narrative; different device from its neighbours (no AT8)" },
    "des_07": { "caption": "Argentina lead, but the gap is one upset wide", "subtitle": "Champion probability, 100k Monte-Carlo runs, as of 2026-06-18", "backs": "ana_01",
                "rationale": "takeaway-title (conclusion, <10 words, active) + descriptive subtitle (metric/method/date); not 'championship probabilities'" }
  }
}

• masthead.headline / standfirst / kicker — the three masthead strings the Programmer renders verbatim. headline_device ∈ D1–D16; headline_backs is the ana_xx whose number the headline states (or null for a number-free verdict that still traces to a finding's *direction*).
• items[edt_xx].title — the section title; device ∈ D1–D16; backs the ana_xx/edt_xx.
• items[des_xx].caption (+ optional subtitle for charts/tables) — the figure/photo/table caption; backs the ana_xx/des_xx.
• rationale (every entry) — one line: the device used + which AI-tell it avoids + why the number is honest.

Naming, not editing — the boundary (do not cross it). You write masthead.*, items[*].title, items[*].caption/subtitle, and a rationale per entry — strings. You do NOT add a finding, change a number, introduce a data-* id, reorder anything, or write body prose. If a title needs a number the Analyst never computed, you have over-reached — pick a device that states what the data *does* say (or label the section plainly), never invent the number.

References

• references/schema.json — full copywriter.json structure + field notes.
• ../../frontend-design-pro/references/exemplars/titling_captioning.md — the few-shot GOOD/BAD corpus (T# headlines, S# standfirsts, K# headings, C# captions), each with a Why + cross-ref to PIT-56/57/58. Read it first.
• ../../frontend-design-pro/references/pitfalls.json — the 错题本; PIT-56 (templated headline), PIT-57 (standfirst spoils the hero's reveal), PIT-58 (caption labels an axis, not a finding) are the entries that catch a weak title. The Auditor's check_15_titling_caption_quality greps for them.
• ../references/topic_profile.json — the shared classifier that decides register (sober → plain literal D16; computational → surprising-number D3).
• ../../dataviz-craft/references/annotation_layers.json — pair the chart caption's asserted takeaway with the chart's own annotation that locates the point.
• ../editor/SKILL.md — the Editor's "standfirst primes, never pre-spoils" writing rule, which PIT-57 enforces on your standfirst.

Done when copywriter.json carries a re-titled masthead (headline + standfirst + kicker), a title for every edt_xx, and a caption for every figure/photo/table des_xx — each on a real device, each backs-ed to a real ana_xx/edt_xx/des_xx, none tripping an AI-tell — and you have changed not one finding, number, id, or line of body prose.