From sign-up to a memory that compounds.
About five minutes to your first checked draft. Then the real product starts: everything you approve, reject, or correct makes the next draft closer to yours.
- 1Create your account
Your account is the key.
Create your account and confirm your email — Claude and ChatGPT connect to it directly with OAuth, no key to paste. Only Cursor and the REST API need an API key from API Keys (shown once, so copy it right away).
- 2Connect your AI
Add the Craft connector — Claude, ChatGPT, or Cursor.
Craft is a hosted MCP server, so there's nothing to install. Point your client at it once and the tools are always there.
- In Claude (web or desktop), open Settings → Connectors → Add custom connector.
- Name it Craft and paste the server URL:
https://craft-mcp.dev/mcp
Click Connect — Claude sends you to Craft to sign in with your account, you approve once, and the tools are live. No API key needed. Using Claude Code in the terminal? One command, then
/mcpto sign in:claude mcp add --transport http craft https://craft-mcp.dev/mcp
- 3Seed your voice
Paste writing you trust — or import a whole archive.
Create one profile per voice, then seed it.
craft_seed_texttakes pasted text, single URLs, or your blog's address — it discovers the archive, previews what it would store, and imports complete pieces with their titles and sources attached.your AI client›craft_create_profilename: "My Voice"← profile created · still learning›craft_seed_textfeed_url: "https://yourblog.com"← 14 posts discovered · stored whole, with titles and sourcesnow just ask, in plain language›"Draft my launch post in my voice, then check it."←craft_compose→ draft →craft_check· all clear: yescraft_composebuilds the brief from your real exemplars and rules;craft_checkis the deterministic last-mile gate — no model judges your rules. - 4Teach it your taste
Approve what lands. Reject what doesn't. It compounds.
The memory is built from your verdicts, not from vibes — and every teaching action is visible and reversible in the dashboard.
the teaching loop›craft_approve"keep this — direct, no hype"← stored with your rationale — future briefs cite why it works›craft_feedback"too breathless, cut the superlatives"← rejection stored · suggests the hard rule it found: ban "game-changing"›craft_learn_from_editdraft + your edited version← learns from the diff — what you cut, what you keptSoft guidance becomes hard rules when you say so:
craft_add_ruleturns "never use em-dashes" into an enforced check with a tier — style rules yield to intentional craft, policy rules never do. - 5Put the toolkit to work
An editorial desk, not just a checker.
Once a profile knows your voice, the rest of the toolkit opens up — all deterministic, all yours:
- Pick the best take.
craft_rankscores candidate drafts — a rule-breaking draft never beats a clean one.craft_reviserecovers a failing draft with accountable, grammar-safe fixes and a redline of every edit. - Arm the guards.
craft_enable_guardswitches on the anti-slop pack, your opener/closer signature, the learned diction band, and quote/attribution integrity.craft_allow_usagedismisses a false positive into exception memory — the same usage never nags again. - Keep house terminology.
craft_glossarybinds deprecated terms to the sanctioned ones ("log on" → "sign in"), fixed by replacement, never deletion. - Teach it what you believe.
craft_positionstores your declared positions and lenses; relevant ones reach the brief, and checks note when a draft touches one — advisory, never a violation. - Audit everything. The dashboard shows every memory with its date and provenance — remove any of them — plus the full teaching timeline, and
craft_style_guideexports the whole voice as portable markdown you own.
- Pick the best take.
The full playbook, one page per move.
The five steps above get you running. These four pages make you fluent — the workflow, the enforcement, and the patterns for each tool you write in.
Your first profile
From a connected AI to a seeded voice — paste, import an archive, and read what it learned.
→2The writing loop
The daily rhythm: brief, draft, check, teach. Plus picking the best take and recovering a failing draft.
→3Rules & guards
Bans that actually stick, guards you can arm, house terminology, and how to dismiss a false positive.
→4Craft in your tools
What the workflow feels like in Claude, ChatGPT, and Cursor — and how to wire the API into CI.
→Run the loop on your own writing.
Seed a profile with work you trust, and the next brief already sounds like you. Everything after that compounds.