A local browser app that takes a business or a personal brand from one interview to a complete brand pack: strategy, logo suite, colour and type systems, brand voice, social kit, AI prompt libraries, design tokens, a video kit and a guidelines document. Your machine, your name on it, your keys.
What you build
| Part | What it is | What it does |
|---|---|---|
| Brain | Claude Code, then one text model through OpenRouter or Ollama | Claude builds the app. The runtime model drafts the written half of a brand - strategy, personas, voice, taglines, prompt libraries - and critiques its own draft. It never does arithmetic. |
| Memory | One folder per brand on your own disk | brand.json with a status and a source on every field, assets in twenty numbered folders, an append-only history file. Every change is a proposed diff, applied only when you click Approve. |
| The keys | Settings, white-label and a local key vault | Your studio name, logo and accent colour. Your own keys for text, images, voice and video: stored on this machine, masked on screen, never logged, never exported. A keyless demo adapter for each. |
| The blocks | Ten blocks of work, one screen each | Strategy, verbal identity, colour and type with computed contrast, the eleven-variant logo suite, visual language, UI system, social kit, print and sales, the AI brand system, a completeness pass. |
| Eyes | The brand home | Ten block cards with counted completeness, what is still draft, what is missing, and what you have spent - every number computed from the store, none estimated. |
| Hands | The exporter | Assembles the numbered folder, enforces the naming convention in code, writes the guidelines and cheat sheet, refuses any draft field or key-shaped string, and zips it with a provenance manifest. |
Where this comes from
A written specification, Complete Brand Pack: a detailed guide, lays out what a brand pack has to contain before the word "complete" means anything: fifty numbered areas, from brand foundation and archetype through the logo suite, colour and type systems, visual language, UI components, social sizes, stationery, the AI prompt libraries, a video kit, accessibility rules, a naming convention and a guidelines document. The specification was written for this lesson and is not published, so there is no link. What the pack takes from it is the shape: the ten blocks it collapses the fifty areas into, the twenty numbered folders, the file-naming rule and the design-token names. Everything else here is original, and the two example brands, a bakery and a fractional CFO, are invented.
Most of a brand pack is the same work every time. The interview changes, the judgement changes, and the rest is assembly: ramps computed from a hex, a type scale computed from a base and a ratio, eleven logo variants composed from one mark, profile images at the size each network asks for. So the lesson does not give you a brand pack. It gives you nineteen prompts that build the app that makes brand packs, and the app is yours, running on your laptop, under your own name, with your own API keys.
What you build
When it runs, it feels like an agency's process with one person in it. You answer twelve questions about a business. The model drafts a strategy and then critiques its own draft: which lines could be pasted on a competitor's site, which contradict each other, which claim something the interview never said. Every field carries a chip that says draft or approved, and where it came from. You approve one at a time, because there is no Approve All button anywhere in the app. Then the same loop runs through voice, palette, type, the logo brief and the marks it generates, the icon set drawn in code, patterns that tile without a seam, twenty social templates, the deck, the stationery, and a video kit built as HTML animation in the brand's real typeface.
The thing it deliberately does not do is arithmetic with a language model. Contrast ratios, WCAG results, tints, the type scale, CMYK conversions, file names, folder assembly and the manifest are all code, and a palette with a failing contrast pair cannot be approved by anyone. It also does not post anywhere, does not hold anyone's Figma token, does not draw true vector logos, and does not clear a trademark. It says each of those in the guidelines it exports.
Do you need this? Be honest first
If you make one brand a year, for yourself, use a designer and a notebook. This is for the person who makes brands for other people: the freelancer who delivers a logo and a colour list and knows a client expects more, the one-person agency that wants a repeatable process instead of a folder called final-final, and the builder who wants to hand a client a system rather than a seat on somebody's platform. It is also a good lesson in the human-in-the-loop pattern, because every screen is one.
You are building a drafting tool, not a designer. The judgement stays yours, the client's name never goes into an image prompt, and an AI-generated photograph is labelled as a reference for direction, not as stock the client may publish. If you are not comfortable saying that to a client, close the tab.
What you need
Any laptop. This build is the exception in the series: the app binds to 127.0.0.1, holds the brand files on your own disk, and there is nothing to keep awake at night.
| Option | Good for | Runtime model |
|---|---|---|
| Windows with WSL2, Mac with Apple Silicon, or Linux with Omarchy | Building and using it | OpenRouter free models |
| A Mac mini with 64 GB or more | Keeping a client's unreleased identity in the building | Ollama with qwen3:32b or llama3.3:70b |
| A small VPS | A team sharing one builder, after you add authentication yourself | OpenRouter |
Every option needs Claude Code logged in (Pro, Max or Team), Node and git. No Docker and no database server: one brand is one folder. The setup guide assumes none of it is installed and starts at "open a terminal". Images, voice and video each have a keyless demo adapter, so all nineteen prompts run for $0; when you want real output you add your own key for the provider you pick: OpenRouter, Google Gemini or Higgsfield for images, ElevenLabs for voice, Replicate or HeyGen for avatar video. Nothing of the author's is in the pack. Every key is yours.
The setup guide is one HTML file you open in a browser. Every command has a copy button, every step has a checkbox, and a progress bar shows how far along you are.
What is in the zip
brand-pack-studio/
README.md the document: what, why, hosting, the accounts, the folder map
index.html the clickable setup guide with all nineteen prompts (same as the browser link above)
COSTS.md every cost, with the date it was read
CREDITS.md where the idea comes from, the as-is notice, what was verified, terms
paths/ one README per platform: Windows WSL2, Mac Apple Silicon, Linux Omarchy
prompts/ 00-project-brief ... 18-completeness-pass
assets/ the cover art
What was verified and what was not
This pack is distributed as is, as a proof of concept. The setup guide was opened in a real Chromium and driven by a script on 2026-09-06: the three platform tabs and their step counts, the progress bar, the checkboxes, progress saved across a reload, the copy buttons for commands and prompts, the prompt text matching its source file verbatim, and the fallback copy path when the file is opened straight from disk. Twenty-five checks, all passing. Every text file was scanned for keys, private addresses and mailboxes before the zip was built.
The first eighteen prompts were written against a reviewed design and read twice by a separate model (Codex), read-only, for contradictions and missing steps: twenty-one findings on the first pass, fourteen on the second, every one triaged and its fate recorded in the pack's review log. Prompt 18, the completeness pass, was added after an audit of the first eighteen against the fifty-section specification and was read once by the same model. The prompts have not been run as a sequence on a fresh machine, the three setup paths have not been executed end to end on three fresh machines, and no provider adapter has been called against its live API. Every prompt ends with a "how you know it worked" check you can see without reading code, and the last two prompts make Claude break the guards on purpose to prove the tests can fail. If a prompt breaks on your machine, paste the full error into Claude Code and ask it to fix it, then post what broke and the platform in the community.
Read the prompt before you run it
Hand each prompt to Claude first and ask what it plans to build. Compare that to what the prompt says. A prompt costs a minute to fix. A wrong build costs an afternoon.
