Flows beat prompts: how one person's know-how becomes everyone's

Every company we talk to has a version of the same artifact: a shared doc, a Notion page, a pinned Slack message, titled something like "Good prompts for X." Someone brilliant wrote it after figuring out the right way to get AI to review a contract, or draft a reply, or reconcile a ledger. It was great advice. Nobody uses it.

Not because people are lazy. Because a prompt you have to remember, find, and paste is a prompt that loses to the three seconds it takes to just type the request straight into the chat box and hope for the best. Good practice that lives in a document is good practice that only the person who wrote it actually benefits from.

A flow is a prompt with a job

A flow isn't a better-written prompt. It's an instruction that runs automatically, attached to the task rather than to the words someone happened to type. When Marcus on Finance types "reconcile October bank statement with Xero," he doesn't need to know that reconciliation should match every line, explain every variance, and draft the adjusting entries in a specific format. That's already decided. It's in the flow. He just gets the output.

This is the difference between know-how and knowledge. Knowledge is written down somewhere. Know-how is what happens automatically when the right person does the task. Flows turn the second kind (the kind that usually lives only in one person's head) into something that runs for everyone, on every desk, the moment they ask.

What's inside a flow

Four things, and they're all visible to whoever builds it, never to the person running it: a model suited to the task, an effort level that matches how much thinking it actually needs, a playbook (the way-of-working instructions, like "Reviewer: find problems, never approve alone"), and, for anything with real steps, a breakdown into subagents so a big task gets planned, executed, and checked rather than answered in one breath.

And when your AI team improves the flow (swaps the model, tightens the playbook, adds a subagent that caught 31% more issues in testing), that change reaches everyone the next time they open their tool. No retraining, no re-pasting, no "did you see the updated doc." One person's judgment becomes the whole company's default, automatically.