Voice JSON

Voice JSON is the minimal structured payload that turns every layer of your brand voice into training data an AI can actually use. It sits inside system prompts for Claude, GPT, or your internal tools and removes the guesswork that produces hedged corporate copy. The object contains six keys. Voice attributes hold specific terms like pragmatic but warm or opinionated without being combative instead of the useless friendly. Never say lists every banned phrase with its plain replacement. Tone contexts map how the voice shifts across marketing, error states, onboarding, and support. Example rewrites deliver five concrete before and after pairs pulled from your live copy so the model sees the exact difference between wrong and right. This format arrived in 2025 once teams realized their Notion voice guides meant nothing to machines. It is the executable output after you finish the audit, lock the three layers, and reverse engineer your highest confidence copy from the founders first email or that one landing page that just felt right. Mailchimp used an early version to keep their warm pragmatic voice intact when they shipped new automation flows in 2024. The JSON forced every error message to stay useful first and warm second without sliding into legal speak.

What it is not is a replacement for the full voice guide humans read. It contains zero theory about how voice stays fixed while tone reads the room. You will not find the voxel diagram, the five brand teardowns, or the long Innocent Drinks carton analysis inside it. Voice JSON also is not a vibe document or a workshop output. If it includes generic attributes like authentic or innovative it has already failed. It is not the 40 page PDF that nobody opens and it is not a static artifact. You prune the never say list every quarter when new jargon like circle back or low hanging fruit creeps in. Most of all it is not magic. Feed it a voice your team never actually uses and the outputs will still feel off. The format only works after you have done the hard work of collecting 20 live pieces of copy and identifying where your current writing drifts into different personalities across surfaces.

Concrete example. Here is the Voice JSON built for Hey in 2026 based on their opinionated anti corporate positioning. It captures the voice that indicts surveillance email instead of describing features.

{ "brand": "Hey", "voice_attributes": ["opinionated", "anti-corporate", "blunt", "principle-driven"], "never_say": ["leverage", "ecosystem", "delight", "seamless", "empower", "synergy", "utilize"], "always_prefer": ["name the enemy", "short declarative sentences", "invite shared outrage", "ground every claim in a specific example like tracking pixels"], "tone_contexts": { "marketing": "confrontational but factual, name the broken thing, zero exclamation marks", "error_states": "direct, solution first, no groveling apologies", "onboarding": "assume the user already hates Gmail, reinforce that hatred immediately", "support": "helpful without deference, treat questions as proof the industry failed" }, "example_rewrites": [ { "wrong": "We are delighted to offer a seamless ecosystem that empowers your productivity.", "right": "Email should not spy on you. We removed the pixels." }, { "wrong": "Utilize our platform to maximize team synergy and deliverables.", "right": "Your inbox is a mess because companies like Google designed it that way. We fixed it." }, { "wrong": "Experience the future of email with innovative new features.", "right": "We do not sort your email. You do." } ] }

Paste that exact block into a Claude project and every homepage headline comes out sounding like Basecamp wrote it. The rewrites teach the model the difference between deferential corporate speak and the polarizing conviction that makes people either love Hey or immediately close the tab. A parallel JSON for Innocent Drinks adds keys for the four beat pattern of greeting, fact, joke, human close and bans any word that sounds like a boardroom. Their error state example rewrites a delayed shipment notice into something that admits the van broke down then offers a free smoothie with a small self deprecating joke. These examples are not decoration. They are the actual training data that stops the AI from reverting to 2023 marketing sludge.

Use Voice JSON every time you onboard a new AI tool or expand your team past four writers. Drop it into Cursor rules, custom GPTs, or the system prompt for your 2026 content brief generator. It pays for itself on neglected surfaces like UI strings, billing emails, and 404 pages where voice usually dies. Discord relied on a version of this in 2025 when they added enterprise servers and needed the copy to stay casual inclusive instead of sliding toward corporate distance. Headspace used a rhythm focused variant that included sentence length rules and low frequency vocabulary requirements so the calm deliberate voice stayed intact before the app even loads. The format also solves prompt engineering problems when you chain models together. One JSON keeps the output consistent from first draft to final polish.

Skip Voice JSON if you have not run the full voice audit yet. Without the never say list and the anchor copy the JSON will encode wishful thinking instead of reality and the outputs will feel like a different brand. Do not use it for regulated industries where legal owns every comma. Headspace keeps their version lighter because their voice lives in pacing that needs full paragraphs attached as examples rather than short rewrites. Never deploy a half finished version missing the example_rewrites key. That turns the whole thing back into the vague adjective soup it was designed to kill. If your team still writes every line by hand with no AI involvement this year then the JSON adds zero value.

Voice JSON turns your brand personality from a feeling into training data that survives writer turnover, agency handoffs, and prompt updates.