--- name: ad-copywriting description: >- Write paid ad creative for AssemblyAI products: LinkedIn single-image ads (image headline, proof element, intro text, headline, CTA, destination) and Google responsive search ads, organized by product line and built to three templates (customer result, product benefit, educational offer). Use whenever someone asks for ad copy, ad creatives, paid creative, LinkedIn ads, Google ads, RSA headlines, a campaign brief for Nick/paid ops, or "what should we run for [product]". Also use to review or rewrite existing ads. Encodes Dylan's feature-sell direction, the A2 reference pattern, the four-beat customer-result chain, and every claim and language gate the paid team has adopted. Do NOT use for organic social posts, email, or landing-page copy; use /copywriting or /assemblyai-content for those. --- # Ad copywriting This skill produces paid ad creative that a paid-ops operator can traffic without rewriting. It exists because the September 2026 paid reset found the live ads were generic ("AssemblyAI" awareness), vague on mechanism, and inconsistent with brand and legal language. The benchmark for "good" is the Product Line Ad Set (12 Sept 2026 draft v3); its exemplars are in `references/examples.md`, and the approved proof numbers with their gates are in `references/claims-ledger.md`. Read both reference files before writing. Product facts (prices, language counts, latency statistics, what is included, positioning lines) come from `/product-claims`, the per-product source of truth; the claims ledger here covers what that sheet does not: benchmark figures with named competitors, customer-story mechanisms, educational assets, and the paid-specific gates. Where the two disagree, `/product-claims` wins on product facts and the ledger wins on benchmark sourcing. Read the ledger again before you put any number on an image. ## The five rules that decide everything 1. **Columns are products, not personas.** Every ad belongs to exactly one product line: Async Speech-to-Text, Realtime Speech-to-Text, Voice Agent API, Speech Understanding, LLM Gateway. Never "AssemblyAI" in general, never a persona or solution ("conversation intelligence", "for note-takers"). Dylan Fox: "You don't buy AssemblyAI, you buy a real-time speech-to-text API." 2. **Feature-sell.** The proof is the ad. A benchmark chart with named competitors, a model's latency, a spec grid, a price. Not a value statement. "We have the best multilingual speech-to-text for voice agents. Here's the bar chart. That should be the ad." 3. **One message per image.** A 5 to 12 word headline plus one proof element. Copy adds who it is for, why it matters, and one reason to believe. The landing page carries the rest. (Nick Andrews' template rule.) 4. **Customer results carry the mechanism.** Dependency, failure, named capability, outcome, in that order. A number with our logo next to it is a testimonial, not a result. 5. **Every claim has a source and a gate.** You write the source line and the pre-launch check into the ad spec. If you cannot source a number, it does not go on the image. ## Three templates Each product line gets at least one ad per template. Product-benefit ads follow the reference pattern exactly. ### Product benefit (reference pattern: "A2") One metric, named competitors, one price. The image is a bar chart or a spec grid from a public AssemblyAI page; the footer carries the metric definition, "lower is better", the source URL, and the list price. Copy adds where it runs (languages as breadth, integrations) and where to test it. Reference: Universal-3.5 Pro, "The lowest word error rate on pre-recorded audio", five bars (4.35% vs Voxtral Mini, GPT-4o Transcribe, ElevenLabs Scribe v2, Deepgram Nova-3), $0.21/hr. Full copy in `references/examples.md`. Non-chart treatments (a code sample, a side-by-side transcript) are kept as labeled variants to test against the chart, never as the primary. ### Customer result (the four-beat chain) Pull the mechanism from the case study page itself, never from the headline stat alone. Then write: | Beat | What it is | Where it goes | | --- | --- | --- | | 1. Dependency | What the customer's product had to get right (one phone mic and two voices; a caller reading an address) | First sentence of the intro text | | 2. Failure or constraint | What broke with the previous provider, or what they tested vendors against | Second sentence | | 3. Named capability | The AssemblyAI feature by its feature name: speaker diarization, turn detection, keyterm prompting, PII redaction | Image headline, with the outcome | | 4. Outcome | The customer's number, stated last | Proof element on the image; last sentence of the intro | The image headline is capability plus outcome ("Better speaker diarization cut Siro's support tickets 90%"). The proof element is the number beside something that shows the capability: a diarized snippet, a quote that names the mechanism, or two or three outcome tiles. A quote is used only when it names the mechanism. If the case study does not name the feature that produced the result, the story is not ready for this template. Say so and offer the closest alternative. ### Educational offer Something a person can actually get: a report, a workshop, a guide, a benchmark. The image states the thing and its takeaway. Copy lists two or three concrete things the reader will learn or leave with. "Register" for events, "Learn more" or "Download" for content. Say what they accomplish, never just that the asset exists. ## Ad anatomy (LinkedIn single image) Write every ad as this field list. Nick pastes these fields into Campaign Manager; the design team builds the image from the mock description. | Field | Spec | | --- | --- | | Image | 1200 × 627 px (1.91:1). Product eyebrow, one headline of 5 to 12 words, one proof element, assemblyai.com footer. Chart ads: metric name, "lower/higher is better", source, price in the footer. | | Intro text | First sentence under 150 characters and able to stand alone (LinkedIn truncates around there). Full text may run longer. Customer results put the dependency in that first sentence. | | Headline | Under 70 characters. Product name or customer name plus the outcome or price. | | CTA button | From LinkedIn's fixed list: Learn more, Sign up, Download, Register, Try. Match the destination's real action. | | Destination | Product or content page with a UTM per product column. Must reach a signup path in one click; activation is the primary conversion. | | Proof on image | One sentence describing the visual and why it is the reason to believe. | | Source | The public page or post each number comes from, with date. | | Gate | What must be checked or approved before it runs. Green if none. | | Variants | Up to two alternate image headlines or proof elements, labeled. | Give character counts inline for intro first sentence and headline. ## Google responsive search ad Per product line: three headlines (30 characters or fewer, headline 1 pinned to the product name and price) and two descriptions (90 or fewer). Same claims as the LinkedIn set; no claim appears in Google that is not in the LinkedIn spec. Show the count next to each line. ## Process 1. **Confirm the product line and template.** If the ask is "ads for X", plan one ad per template. If the ask spans products, build a tabbed brief (see Output). 2. **Ground the claims.** Pull product facts from `/product-claims` and benchmark or customer proof from `references/claims-ledger.md`. For anything in neither, use assemblyai.com/benchmarks, /pricing, the product page, or the launch blog, and add the new claim to the ledger with source and date. 3. **For customer results, read the case study page**, not the customers grid. Extract the four beats. Confirm the feature named is the one the ad's product column sells; if not, move the story to the right column. 4. **Write the image first.** Headline, then proof element. If the headline needs more than 12 words to carry capability plus outcome, cut the outcome number from the headline and let the proof element carry it. 5. **Write the fields.** Intro text, headline, CTA, destination, proof, source, gate, variants. 6. **Run the language check** below, then the pre-launch gates. 7. **Propose the launch order.** Three or four live creatives per channel at a time. Lead with the ads that need no approvals and prove the feature-sell thesis (product-benefit charts), plus one customer result as a control. ## Language rules (all mandatory) - "Speech-to-Text" is the product word. Never "transcription" as the product or category; "transcript" as the output noun is fine. Case-study titles that say "transcription" are quoted as-is only inside the story link. - Never "production-ready" (banned); avoid "production-grade" in headlines too. - **Language support is breadth, never a single language.** "18 languages with code-switching" (Universal-3.5 Pro and Realtime), "99 languages" (Universal-2). Never a named language as the hook, never "best in every language". Voice Agent API creative does not lead with language at all. - Healthcare or PHI: Business Associate Agreement language only. Never "HIPAA-compliant", "HIPAA-ready", "HIPAA-certified". Run `/hipaa-language-check`. Compliance is "helps meet GDPR, PCI and privacy obligations", never a badge. - Qwen 4B is "optimized and hosted by AssemblyAI", never "custom-trained". Cost claims only versus named external models, never "our cheapest". Never "free model" or "free to use". - Voice Agent API: never "speech-to-speech" as the product name; latency in body copy only, never the headline; no zero-data-retention claims. - Benchmarks: "spikes, not averages". Lead with the dimension we win; never a general leaderboard position as the whole ad. Date-stamp any live-board snapshot (Coval reruns every 30 minutes). Pair latency with the condition ("on real agent conversations"), never a bare speed number. - SLA is 99.9% if mentioned. No "we never train on your data". No published discount percentages. - No em-dashes in ad copy. Plain, concrete sentences. State what the API does. Cleverness, if any, lives in the headline only. - Never "conversation intelligence" or other solution categories as the frame. - Model names spelled out and hyphenated as the pricing page has them: Universal-3.5 Pro, Universal-3.5 Pro Realtime, Universal-3 Pro Streaming (when quoting a customer who deployed it), Qwen3.5 4B. ## Pre-launch gates (write them into every spec) - A named competitor comparison for the Voice Agent API needs Dan's sign-off. - Any number from a live third-party board is re-checked the day before launch and dated on the image. - A customer story runs in a product column only if the case study names that product's feature. Otherwise it moves. - Beta-labeled features (check docs) and features whose price disagrees across pages do not run until settled. - Deprecating features: exclude legacy accounts until the deprecation notice is out. - Events: confirm date, format, and registration URL; geo-target to the metro. - Audiences: exclude AssemblyAI employees; US, UK, CA only until localized creative exists. ## Per-product leads (what each column sells first) | Column | Lead with | Best public proof | Known gap | | --- | --- | --- | --- | | Async Speech-to-Text | Accuracy on real audio; code-switching across 18 languages | /benchmarks WER, entity, diarization charts; Coval #1 of 31; Siro, Calabrio, JotPsych stories | None; the strongest column | | Realtime Speech-to-Text | Entity accuracy in real time; turn detection; 18 languages | /benchmarks realtime missed-entity chart; Pipecat frontier; Super story; Goodcall quote | Pipecat figures differ by source; pick one | | Voice Agent API | One WebSocket; $4.50/hr all-in; alphanumeric accuracy vs OpenAI and Deepgram; working agent in an afternoon | Product-page comparison table; live demo; the RX-7704132 transcript | No named public customer yet; CBB Labs pending confirmation | | Speech Understanding | PII Redaction (about half of spend), Speaker Identification on the diarization chart, then Summarization, Sentiment, Entity Detection | Guardrails comparison grid; diarization cpWER chart; Calabrio GDPR quote | Summarization is beta-labeled with a price conflict | | LLM Gateway | The specific model and its proof (Qwen 4B: ~600 ms, 94% cheaper per audio hour than GPT-4.1), then VoiceBench | Qwen launch post; /products/llm-gateway demo; VoiceBench results page after 28 Sept | No public customer story; own-product proof (Dictation API pipeline) until one is approved | ## Output - **Single product, one or two ads:** the field list per ad, in Markdown, plus the Google RSA block and the gates. - **Multi-product set or a brief for review:** an HTML brief on the brand design system (Cobolt accent, serif display, mono labels) with an Overview tab (feedback rules, anatomy, launch order, verification list, sources) and one tab per product line. Each ad renders as a 1.91:1 image mock beside its field list. Publish as an artifact so reviewers get one link. The Product Line Ad Set is the template for this. - Always end with the verification list: what must be true before each ad goes live, and who owns the check. ## What not to do - Do not write a headline that would work for any speech vendor. - Do not put two messages, or two charts, on one image. - Do not lead a customer result with the number and our logo. - Do not use a persona or use case as the ad's frame. - Do not quote a customer saying we are "great"; quote them naming the feature. - Do not put latency in a Voice Agent API headline, or language in any Voice Agent API copy. - Do not average benchmarks or cite a general leaderboard rank as the whole ad. - Do not ship anything whose source line you could not write. ## Related skills `/product-claims` for every product fact; `/knowledge-base` or `assemblyai-product-facts` for wider grounding; `/hipaa-language-check` for anything healthcare; `/medical-mode-claims` for Medical Mode numbers; `/customer-quote-priority` to choose which customer leads; `/supporting-imagery` (Graphs style) to build the chart the ad shows; `/copywriting` for non-paid channels.