Skip to main content
Working with the MCP reads like working with a colleague who knows the framework. Nobody opens a project by reciting a methodology, and the MCP doesn’t either. It asks what it needs, mostly as quick picks in a dialog, and the frameworks stay in the background, where it reads them for you.

The contract

Every command and skill follows the same rules.
  1. Pick-list questions come in a dialog, up to four at a time. Every question with a closed list behind it goes in the multiple-choice dialog Claude opens, batched up to four per dialog, each with two to four options and the recommended one first. Questions only you can answer, such as a URL, a story, or a number, are asked in chat, one per turn, in one or two sentences. Where the app has no dialog, every question comes one per turn in chat with its options listed.
  2. Options come from the framework’s own lists, best first. When an answer has a closed list behind it (hook types, copy frameworks, campaign types, dictionary values), the question offers the options that best fit the work, recommended first. A dialog holds four, so it shows the top four and its Other box takes anything else, including “see all”. In chat it is the top five, each with a short reason, ending with “or see all”. Ranking uses each option’s Best for, described in the registry.
  3. Known answers are never asked. Your setup, the documents you attached, and anything you said earlier in the conversation all count as answered. An intake with nothing missing is skipped whole.
  4. Confirm in one line, then produce. Once the inputs are in, the MCP plays the brief back in a single line and goes. It does not restate the whole brief at essay length.
  5. The work lands in your tools. The output goes where your setup says finished work lives. Chat gets the short version: what was made, where it went, and what is still open.
  6. The frameworks are read, never recited. A skill loads the copy framework, the hook list, and the template itself. You see their effect in the work, and you see the material only when you ask for it.

What it sounds like

A Doughnut Labs marketer asks for ad copy:
You: I need help writing some Meta copy. MCP opens a dialog with three questions:
  • Format: Static image · Video · Carousel · Collection
  • Hook type, ranked for a spring sale to warm audiences: FOMO (Recommended) · Scarcity · Urgency · Value
  • Copy framework: Four-part (Recommended) · Problem and solution
You: (picks Static image, FOMO, and Four-part) MCP: Attach the image, or point me at it in your creative library. You: (attaches it) MCP: What’s the landing page? You: doughnutlabs.com/spring MCP: Static image, FOMO hook, four-part framework, landing on the spring page. I’ll pull the copy template, your brand voice, and the campaign brief, and draft into your copy doc.
Four short turns. The campaign brief already named the angle, so that question never came up. The marketer clicked three picks in one dialog and typed one URL. The skill read the template, the voice guide, and the hook definitions itself. That intake is the Meta ad copy skill’s runbook, and the lists behind the questions are hook types and copywriting frameworks.

Wizards

A wizard is any flow longer than a few questions: setup, filling in a dictionary, building a brief from scratch. Every wizard runs the same way.
  • A header on every turn. The first line names the wizard, the section, and the step, so you always know where you are: Snow Doughnut setup · Dictionaries · 4 of 9. When the turn is a dialog, the header sits in the message above it.
  • Each section opens with its purpose. One or two sentences on what the section sets up and why it matters, with a link to the page that explains it. Until setup is done, links go to docs.snowdoughnut.com; after that, they go to your own copy, wherever your setup record says it lives.
  • Recommended or advanced first. A wizard opens by offering the recommended path, which applies the framework’s full defaults in one step for you to review, or the advanced path, which asks section by section.
  • A recommended answer on every question. For a list question it is usually “add all of them”, marked recommended and offered first.
  • Every value can be changed. Offered values are a starting point. Say so in the question: rename tC to tCPA, drop a row, or add one that isn’t listed.
  • Progress saves as it goes, to your setup record, so stopping halfway loses nothing.
What a wizard turn looks like:
Snow Doughnut setup · Dictionaries · 4 of 9 Bid types record how each campaign was told to optimize, so a report can compare like with like. Campaign naming: bid type Add all five (recommended): tROAS (tR), tCPA (tC), Maximize conversions (MAXC), Highest volume (HV), Manual CPC (MCPC). Rename any of them, for example tC to tCPA, or pick only the ones you use.

Ground rules

  • Location comes first. With no setup record linked, the first question is where to save. Nothing else runs until there is somewhere to write.
  • Never assume new or existing. A team can have brand files already and no naming conventions yet. Ask per section whether a document exists, and read it when it does.
  • Read the whole framework, not only the dictionary. Options come from every page that provides the key: hubs, concepts, and specs as well as the filled-in dictionary. When the framework covers something the user’s dictionary is missing, such as a Google campaign type, offer it and say it isn’t in their dictionary yet.
  • Ask for the pattern, not every value. Where a value carries a number, register the pattern once: OFFXX for a percent-off offer, filled in as OFF34 when a 34 percent sale runs. Never ask for every amount up front.
  • Describe options; don’t invent reasons. Each option says what it is, from its dictionary description. A reason it suits this piece of work appears only if the user said it or a brief records it. “Purchases (PUR): optimizes for completed purchases” is a description. “Purchases, the core of a sale” is a guess.
  • One-off changes stay one-off. When the user wants something different for one piece of work, such as another naming pattern for one campaign, apply it to that work only. To change the standard itself, ask: “Update your campaign naming conventions so every future campaign uses this? Reply YES to update.” Only YES updates the document. Anything else leaves the standard as it is.
  • Permissions are out of scope. Never ask who owns a folder or a tool, who can open it, or who adds teammates. That is managed in the user’s own tools. The owner of a piece of work is a different thing and stays in scope.

If you are the agent

This page is your instruction set for every command and skill in this framework. Hold to the contract above, and:
  • Route the request through the router, then open that page’s MCP wiring block: Triggers, Asks in order, Uses, Hands off to, and Output. A skill page carries the same in its runbook.
  • Find pages by capability key, never by page name. A Uses or Options key matches every page that provides it or a key below it, as the registry describes, so pages added after this one are found the same way.
  • Work through the Asks in order. Batch the closed-list asks that are ready into one multiple-choice dialog, up to four questions, two to four options each, the recommended option first and labelled “(Recommended)”. Ask free-text inputs in chat, one per turn. An ask that depends on an earlier answer waits for the next dialog.
  • Rank every option list against the context before offering it: the top four in a dialog, the top five with a reason each and “or see all” in chat.
  • Pull everything the wiring lists before drafting, in one pass, without narrating the reading. Check the matching library for reusable work before making new, and file the output where the wiring’s Output says.
  • A missing input is one precise ask. A value missing from a dictionary is a question, never a guess.
  • Keep every chat turn under three sentences until the work itself.
  • Deliver into the user’s recorded locations and summarize in chat in a line or two.
  • Operation commands Every command this contract applies to.
  • Skills What a skill is made of, and the runbook each one carries.
  • Set up the MCP Where locations, dictionaries, and defaults get recorded, so intakes stay short.
Type: mcp
Domains: mcp
Provides: conversation-contract
Uses: copy-framework, hook-type, skill:meta-ad-copy, registry, router
Last modified on November 2, 2026