Prompt Anatomy

Prompt anatomy is the mandatory five block structure every prompt needs if you expect it to survive more than one model update and actually get used by more than one designer. The blocks run in fixed order. System sets the role with precision. You are a senior brand strategist at Linear who owned the 2025 visual identity system and rejected 187 hero concepts before finding the right ones. Scope draws the exact boundary. Evaluate only the hero headline subhead and primary button copy against the brand rubric. Do not address layout color or illustration. Examples follow and they carry the real weight. You insert four real cases from your last two quarters two that the team approved and shipped two that got sent back with specific notes about voice drift and clarity failures. The model learns the taste by seeing the taste not by reading your instructions about it. Constraints come next and they act like guardrails on a mountain road. Never use em dashes. Never begin with Imagine a world where. Never propose copy longer than nine words. Never invent product benefits not listed in the brief. Output format finishes the prompt and turns it into an API call. Always respond with this exact JSON schema containing overallScore voiceMatch clarityRating issues array and suggestedRevisions array. No extra prose outside the JSON or the downstream eval pipeline breaks. This anatomy is what separates teams that treat prompts as disposable strings from teams that treat them as durable assets in 2026. The system role prevents tone drift when the base model changes its default personality. The scope stops the model from confidently answering questions it was never hired for. The examples replace vague adjectives with pixel level truth. The constraints eliminate the exact failure modes that waste the most designer hours. The output format lets you plug the prompt into an automated rubric that scores at 11pm so you do not have to. Prompt anatomy is not whatever you happen to type when you open a new chat. It is not a clever hack you saw on a LinkedIn post about prompt engineering. It is not a massive wall of text that tries to do ten jobs at once. It is not the chat history you keep feeding the model because you are too lazy to extract the pattern into a reusable unit. It is not optional the moment your prompt leaves your private workspace. Skip the anatomy and you are back to hoping the model gets it right this time which is exactly how most teams operate and exactly why their output quality collapses every time Anthropic or OpenAI ships something new. The concrete example lives in the ClaudeBrainy pack that started circulating in design teams during Q4 2024. The prompt is called brandHeroReviewer v2.4. Its system block reads You are a principal voice designer who shaped Mailchimp voice guidelines from 2022 through 2025 and trained three junior writers who now run brand at high growth startups. Scope limits it to hero elements only on landing pages no blog posts no email campaigns. The examples section contains five paired cases. Each pair shows the original copy the team feedback the final approved version and the rubric scores. One failure case involved a headline that scored low on clarity because it buried the lead. The constraints block is merciless. No sentences longer than 18 words. No questions used as headlines. No exclamation points unless celebrating a concrete metric. No corporate jargon like revolutionize or ecosystem. The output format requires a fixed YAML structure with fields score toneAlignment issues list revisions list and confidence. This exact prompt runs in three variants. The fast variant drops one example and returns lighter JSON for IDE use inside Cursor. The strict variant adds two more constraints for final ship reviews. The author variant flips the system role so the model generates instead of critiques. All three share the same spine and all three are versioned in the git repo with evals run against a 60 case test suite before every merge. When the team migrated from Claude 3 Opus to Claude 4 Sonnet the prompt needed only a patch release because the anatomy kept the output stable. The Ramp design team learned this the hard way in March 2025. Their unanatomized expense policy audit prompt worked in February then silently degraded across four different Slack copies and Cursor configs after an OpenAI update. Nine days of bad audits went out before anyone traced the drift. That cleanup cost real velocity. Deploy this anatomy any time a prompt has been used twice or shared once. Deploy it before you add the prompt to your team library whether that library lives in Claude Skills Cursor rules Continue dev config or OpenAI prompt management. Deploy it when you are composing larger systems where a parent audit prompt calls this hero reviewer as one of its children. The anatomy becomes non negotiable the moment the prompt touches anything customer facing or anything that feeds your design system. Skip the full anatomy during pure discovery sessions where you are stress testing a new model on random tasks. Skip it for personal one off ideas that will never be versioned or evaluated. Skip it at 3am when you are exploring wild concepts and the goal is velocity not consistency. Those moments reward flexibility over structure. The teams moving fastest in 2026 built their prompt libraries the same way they built their component libraries in 2018 with strict anatomy versioning and ownership instead of Notion pages full of rotting strings.