From an empty folder to GitHub: how AI Consulting was built in Flow

What a Playbook is

A Flow Playbook is a folder of living documents that carries an expert’s method. It holds what they know, the sources behind it, worked examples, and tools a reader can use on their own work. Each file is plain Markdown. The folder carries its author, a version and a licence. You build it in Flow with any AI you connect, you review every change, and you publish it to GitHub. Anyone can bring it into their own Flow with its address, and when you publish an update, it reaches them as a change they can review.

AI Consulting is the first Playbook, and it launches with Flow. Its author is Manav Sehgal, who spent nine years at Amazon leading product, solutions and business development across AWS, Alexa AI and Shopping. The Playbook introduces Amazon’s mechanisms, then applies Working Backwards and the PR/FAQ to adopting AI, from picking the right project to taking a pilot to production. This is how it was made, all of it in Flow.

The answer first

1. Add a folder and choose it. Chat works with no document open. It answers from the folder and writes into it.
2. Ask Chat for a front page. It writes the README in your voice and marks every claim you still need to check.
3. Bring in what you already have. New from Web Page turns a shared AI chat or a web page into a document, links kept.
4. Let Ideas pick a Skill. Press Suggest, then run a card. A figure, a form or a table waits in Review Changes until you keep it.
5. Ask Chat for the next document, with your own pages as sources. Repeat until the path is done.
6. Publish ▸ GitHub Repository. Name the repository and the folder. Flow sends the files and nothing else.
7. A reader uses File ▸ New from GitHub… with the folder’s address. When you publish again, their copy shows an update to review.

An empty folder, and Chat already in it

The walk started at 10:55 in the morning on a new Mac identity, with a new folder called ai-consulting-starter and nothing in it. Chat was ready anyway: Answers from ai-consulting-starter · 0 documents. What should we write in ai-consulting-starter? Manav pinned Claude Code › Claude subscription as the route, and every model run in this walk used it.

A front page, in the author’s voice

The first message described the Playbook and asked for a README of about 600 words. Chat wrote 788, with the title, who it is for, what a reader will have at the end, and the path through the documents. Two sentences carried a [check] mark. These were claims Chat could not stand behind on its own, such as the exact wording of a quote and who said it. A mark stays in the document, and the to-check badge counts it, until the author answers it or resolves it.

An old AI chat became a reference

Manav had already asked ChatGPT to list Amazon’s mechanisms with a source for each. He brought that shared chat in with New from Web Page: the table, all 24 rows and 49 links came across as a document.

Then he asked Chat to restore the links, restructure the document, rename it Amazon Mechanisms and Practices, add an executive summary picking the mechanisms that matter for this Playbook, and update the README to point at it. Three minutes later, Review held 17 changes across the two files. Chat said plainly that it could not load the shared page, so it found sources itself: 14 public pages. It marked the six mechanisms it could not source with [check] instead of guessing.

Ideas picked a Skill for a section

Pressing Suggest in Ideas read the README’s sections and named a Skill for two of them. It offered a Visual Explainer for Amazon Mechanisms, briefly and a Guided Exercise for How to use this Playbook in Flow, each with a sentence on why it fits.

The Visual Explainer’s sheet showed what it would read (279 words), its five steps and its cost: Included in a subscription you already pay for. About a minute after Run Skill, the figure was waiting in Review, right after the section it read: Good intentions don’t work; mechanisms do, built in four steps a reader can play. Manav kept it.

The next document, from his own pages

Next, Manav asked Chat to read his LinkedIn profile, his newsletter, his Substack and his story on orionfold.com, and to build the Mechanisms Primer with Flow Elements. In under two and a half minutes Chat created it. The document had 1,624 words with a checklist of symptoms a reader can tick, tables, a stepped figure called Which part is missing?, and a Try it section. It linked the README to it in two places, and it said what it found on his pages and what it did not. It found one real career example and Alexa Conversations growing from one language to nine, and no post of his on mechanisms yet. Three claims went in with [check].

The same pattern made the PR/FAQ document, 1,060 words. Chat then asked two questions back instead of deciding for him. What should it be called, and where does it sit on the path?

Figures you can press and step through

An Isometric Figure drew the path through the Playbook as a stack of glass layers, one per document. Press a layer and it lifts, with a note on what that document does.

On the PR/FAQ document, Ideas made four suggestions. A Visual Explainer on Run the review, a 75-word section, came back in under 30 seconds. Its steps were read in silence, questions in order, one of three decisions, then write the decision down.

Review changed the plan, and the source followed

The README’s first path had six documents. Reading them back, Manav cut it to four. They are the Mechanisms Primer, Amazon Mechanisms and Practices, the PR/FAQ for an AI project, and Pilot to Scale. He asked Chat to bring every document in line with the new path, and each change waited in Review. A sixth file, Flow PR-FAQ, is a worked example of the method. It is the PR/FAQ for Flow itself.

Publishing is yours, not Chat’s

When Manav asked Chat to publish the folder to GitHub, it said no: I can’t publish from Chat. I can only change documents in this folder. Then it listed what to look at first. The repository is public, a date needs correcting, a few sentences are still in the future tense, and some entries are still marked to check.

So he did it himself. Publish ▸ GitHub Repository, destination orionfold/flow/flow-domains/ai-consulting-starter, then Publish to GitHub: 6 files sent. The commit added the six Markdown files, byte for byte, and touched nothing else in the repository. The only setup was a GitHub token, pasted once.

The same folder can go out as a GitHub Pages site, a PDF, Word, Excel, PowerPoint or an EPUB. The preview for GitHub Pages plays the figure step by step. Flow’s own interactive features, such as Chat on the Playbook, stay in Flow; an exported file is for reading.

A reader brings it in, and stays current

Into a second folder, File ▸ New from GitHub… took the folder’s address and read it. The sheet showed orionfold/flow · flow-domains/ai-consulting-starter · 6 files. The six files arrived byte for byte. The sheet also said No licence stated, so Manav chose one, CC BY 4.0, added it to the README and published again.

The reader’s copy noticed. Its folder showed a blue dot. An update was available.

Pressing it listed what changed: README · Updated. You have not changed this one, so replacing it loses nothing. The sheet checks each file against the reader’s copy before anything is replaced, and the reader picks what comes in. Update Selected brought the new README in.

Last, we tried it the way a stranger would. We used a brand-new Flow on a first launch, holding only the bundled Guide. New from GitHub read 6 files, 75 KB, the version, and Licence: CC BY 4.0. After Add, the README opened by itself and its figure played all four steps. No model ran and nothing was spent.

The numbers

Your agents, on your Playbooks

With Flow MCP, the Claude Code or Codex you run in Terminal can use your Playbooks too. In Settings ▸ General ▸ Your agents, open Flow to them as Read or Read and propose. An agent can read a Playbook’s summary and reading order, read and search its documents, and apply its method to the work it does for you. Whatever it proposes comes back to Review Changes under its name, and Flow keeps a log of each call.

Try it

Get AI Consulting on GitHub. In Flow, choose File ▸ New from GitHub…, paste the folder’s address and press Add. Bringing a Playbook in is free on Base; Chat on it needs Flow Pro. Start with the Mechanisms Primer, pick one real customer task, and work along the path to a PR/FAQ for it.

If your field has a method of its own, the blueprints on the Playbooks page are yours to take. Each is a plan for a Playbook in one field, built the way this one was. Build it, put your name on it, and launch it as your Playbook.