How it works

From brief to build context

Scaffolds structures the website, people and AI contribute the reasoning, and what comes out the other end is a portable project model rather than a folder of documents that disagree with each other.

The sequence

Eight steps, in order

  1. 01

    Start with intent

    Name the audiences and what each is trying to achieve, before any page exists. Structure is downstream of intent, so deciding it first is what stops a sitemap becoming a list of departments.

  2. 02

    Shape the architecture

    Build the page tree and the blocks each page is made of. Record the purpose of every page and section as you go, and raise anything unresolved as a red flag rather than quietly deciding it.

  3. 03

    Map the journeys

    Turn each intent into an explicit path across the architecture and into external systems, with an entry, a goal, ordered steps and an exit.

  4. 04

    Connect your AI

    Pick Claude, ChatGPT, Gemini or your editor and follow that product's own few steps. It then reads the goals, the rules, what already exists and what has been agreed, without anybody exporting or pasting anything.

  5. 05

    Or take the file, if yours cannot connect

    The whole project exports as one Markdown file carrying its own editing rules, so a provider with no connection and an organisation that will not allow one both still work.

  6. 06

    Answer what it will not decide

    When a connected AI reaches something nobody has settled, it asks you rather than guessing, and the answer is recorded against the project instead of living in one chat. A decision you have already refused is not put to you twice.

  7. 07

    Review every change

    Whether a file came back from a chat or a proposal arrived from a connected AI, it is parsed deterministically and every addition, edit, move and removal is listed before anything is written. Removals are counted separately.

  8. 08

    Hand the same context forward

    The approved structure, its reasoning and its journeys travel into delivery as one readable model rather than dying in a slide deck.

The part people ask about

Can the AI change my project without asking?

No, and this is the single most important thing on the page. A returned file is parsed deterministically, with no model anywhere in the reading of it. What you get is a change list, not a changed project.

Removals are counted and confirmed separately, because they are the destructive ones. Applying is a single revision, so undo puts everything back.

Try it below. Nothing on this page talks to a real project.

Review changes

5 changes
  • addPage added, Integrations
  • add6 blocks added to Integrations
  • editProduct, note edited on Portable AI context
  • moveEnterprise, Delivery model moved above Roadmap
  • removeHome, Product proof strip removed

Nothing has changed in the project yet.

Working with your own AI

What a connected AI can see, and what it can never touch

A connection is one endpoint that only reads. The second list is not a policy or a setting you could get wrong: there is no method on the other side of that endpoint that does any of it, which is a different kind of promise from one that has to be configured correctly.

What it can read

  • Twelve derived documents rather than one long file, so asking what the rules are does not mean reading every wireframe
  • Every page, block, intent and journey, each carrying the identifier that survives a rename
  • The editing rules, the meaning of every field and the full element vocabulary
  • What has already been settled, including proposals that were turned down, so the same question is not put to you twice
  • Two rules that stop capable models overreaching: scope is authoritative, and assumptions are not facts

What never leaves

  • Every write. It reads, and may ask a question or leave a proposal; changing the project stays a person's act
  • Comments and pinned notes, and the project roster
  • Collaborator email addresses and credentials, which no export and no endpoint carries
  • Access and sharing, which nothing arriving over a connection can grant itself
  • The revision history, which only Scaffolds writes
  • Every other project you have. A connection is made to one project and cannot reach a second

Scaffolds sends nothing anywhere on your behalf. Your own AI comes and reads, or you move the file yourself, so which providers your work reaches stays entirely your decision, and ending a connection is one control, after which the credential is dead.

Questions

The things people ask before they start

What does Scaffolds actually structure?

One document per project holding the page tree, the blocks each page is made of, the audience intents, the journeys across the pages, and the reasoning recorded against each block. It is a planning model, not a design canvas and not a content management system.

How is this different from a sitemap tool?

A sitemap tool records what pages exist. Scaffolds records why they exist. Every block carries its purpose, the evidence behind it, the audiences it serves and anything still unresolved, so the argument survives alongside the structure.

Which AI does Scaffolds use?

None, and that is the point rather than a gap. There is no model inside the product: you bring your own. Connect Claude, ChatGPT, Gemini or your editor and it reads the project where it lives, or export the project as a Markdown file carrying its own editing rules and work on it in whichever provider your organisation has approved. Both roads end in the same review, and changing which AI you use does not touch the project.

Can the AI change my project without asking?

No, and a connected one cannot either. An imported file is parsed deterministically with no model involved, and the complete list of additions, edits, moves and removals is shown before anything is written. Removals are counted separately, and applying the change set is a single revision that undo reverses. An agent connected over MCP reads the project; the most it can send back is a question, or a proposal that waits for somebody to approve it.

Do reviewers need an account?

No, but the owner switches link viewing on first. New projects are private, so until that happens only the people invited to a project can open it. With link viewing on, anyone holding the URL reads the whole project and can leave comments and pinned notes without signing up. Editing is separate and always needs an invitation, because a link never grants it.

What happens to the planning work at handover?

The same structured file goes forward. Page architecture, block sequence, component names, notes and journeys are all readable by a person and parseable by a coding agent, which is what stops the reasoning being lost between planning and build.

What does it cost to work this way?

Nothing for one project. The free plan is a real project with every planning surface on it and your own AI connected to it, rather than a trial that expires. What it does not include is a second project, which is the point at which the work has become recurring and paying for it is honest. Pro and Team are bought with a card, and Enterprise is invoiced against an agreement.

Look at a project that was planned this way

The scaffold for this website is public and read only. It has the six pages, the blocks, the four audience intents and the journeys, with the reasoning recorded against each section.