What happens between the first question and the files.
Every step, in order, with what is a model's work and what is code.
Eighteen questions that ask for stories.
Ask a person how they would describe their voice and they give you a brochure: direct, warm, professional. Ask them about a client they turned down and why, and you get their voice. The interview is built on that difference.
The questions move from easy to revealing. The first three have answers you have given a hundred times. The middle asks for one specific time something happened. The later ones ask what you disagree with, how you open an email, and what you would never say.
Each question has a follow-up. It appears only when a recorded answer runs under thirty seconds, or a typed one under forty words. It asks for the specific: a name, a number, the words you used.
You record one answer at a time, a minute or two each. You can re-record any answer, type instead of speaking, or skip. Spoken answers carry more of your voice, which is why the interview is spoken first.
One answer at a time, uploaded as it finishes.
Your browser records each answer and uploads it when you press stop. A dropped connection loses the answer in progress, never the interview, and you return to the next unanswered question.
Each recording is transcribed by a speech model on Cloudflare's Workers AI. The recording is deleted as soon as the transcript exists, unless you choose on your account page to keep recordings for re-interviews.
The transcript is what the guide is written from. You can read every transcript on the review page before you finish.
The guide is written from your transcript, under a schema, with a check.
A model reads the eighteen transcripts and writes the ten sections in second person. The model is Claude Sonnet 5.5, from Anthropic, under its API terms. The sections and their shapes are fixed by a schema, so every guide has the same parts in the same order.
A check, in code, requires at least two of your own lines in every section, inside quotation marks. Each line must appear in your transcript. A line that appears nowhere is an invented line, and the section fails. The model gets one more try at a failing section. If it still falls short, the guide is delivered with a note on that section.
The three examples in section ten are rewritten by a smaller model with the finished guide as its instructions. The same model does the rewrites on the "try it" page.
Nothing is trained on your words. Your transcript is sent to the model as data, the guide comes back as text, and the transcript stays in your account.
The files, built and checked.
The guide is packaged as a set of files. A Claude Skill folder, checked against the skill format: the frontmatter fields, the name rule, the line limit. A CLAUDE.md snippet that imports the guide. A ChatGPT custom-instructions file under 1,500 characters, plus a project instructions file. A Gemini instruction block. VOICE.md. And a PDF rendered from the guide page. Every file is validated before it is offered, and all of them come as one zip.
Install notes for each file are on the guide page and in the guide to installing a voice guide.
The limits, written down.
A guide is as good as the interview. Short answers make a short guide, and the page tells you when an answer is thin.
An assistant follows a guide the way a new colleague follows a style sheet: well, and not perfectly. Instructions fade as a conversation grows; the files are built so the guide can be reloaded in one line.
The site makes no claim about detectors. The guide makes a draft sound more like you.
Longhand is independent and not affiliated with OpenAI, Anthropic, Google, or Notion. Claude, ChatGPT, Gemini, and Notion are their owners' names and appear in text only, to say which assistant a file is for.
The half-hour version of a longer method.
Longhand comes from a method Brett Wharton uses at NexusMKTG. There, half an hour of his questions and hours of recorded conversation become a dozen brand guides and a trained Claude system. That engagement starts at $8,000 and Brett runs it himself. Longhand is the self-serve version of the first step: half an hour, one guide, the files, and you. If you want the whole suite and someone to operate it, that is the service.