Flow walkthrough, run by an AI

This page is for anyone who just changed a flow and wants it walked end to end before real users hit it. It is one of the ready-made studies in the use cases for visual product feedback hub, alongside AI usability testing.

What a flow walkthrough is

A flow walkthrough is an AI stepping through one journey you name, signup, checkout, inviting a teammate, and reporting what broke, what merely confused, and what could be better, step by step.

It differs from a usability pass in one way that matters: you pick the path. A usability pass decides for itself which jobs a first-time user would try. A flow walkthrough follows the journey you care about right now, which is the right shape when you've changed one thing and want that one thing checked, or when a support ticket says "I couldn't finish signing up" and you want to see it from the outside.

Emails are part of the flow. A verification code that never arrives, a welcome email that links to the wrong page, an invite that renders as one long line on a phone: these are steps of the journey that most reviews skip because there's nowhere to receive them. Cobalt gives the AI a disposable inbox, so they're rendered and judged like any other screen.

How the AI runs it

You describe the flow in a sentence. The AI restates it as numbered steps, says what it will and won't do, and then:

1. Sets up cleanly. A fresh browser session, a throwaway inbox, no leftover logins. If the flow needs an account, it makes its own.
2. Walks the steps. One screenshot per step, at the moment before and after each action where it matters. Where a step forks, it notes the fork and takes the path a real user most likely would.
3. Reads the emails. Any message the flow sends is pulled from the inbox, rendered as HTML, and judged: arrival, appearance, links, and whether the copy matches the page that sent it.
4. Stops at the edges. Before payment, before anything irreversible, before a human-verification challenge. Each of those is captured as a finding, not worked around.
5. Writes the verdict on top. What blocked the goal, what slowed it, what was merely cosmetic, and what worked.

What comes back

A review at a shareable link. The verdict sits at the top, the steps below it in order, each with its screenshot, what happened, and any finding. Findings are ranked by outcome for the person walking the flow:

• Blocked the goal. The step cannot be completed, or the user would abandon.
• Slowed the goal. Completed, but with a wrong turn, a retry, or a moment of doubt.
• Cosmetic. Noticed, not costing anyone the outcome.

Export it as markdown for a coding agent, a PDF for a teammate, or push it to GitHub as an issue. Run it again after the fix and the review leads with what changed.

Running it

This study needs a browser, so it runs from a coding agent. In Claude Code:

/mcp__cobalt__flow_pass https://example.com

Then describe the flow when it asks, or put it in the same message: "sign up, create a project, then invite a teammate." The studies page has the paste-ready block for Cursor and Codex.

From Claude chat, it covers the public, logged-out part of the flow and says so before starting.

When to run it

• After changing a flow, on staging, before it ships.
• When a support ticket describes a journey failing and you want it reproduced from the outside.
• After changing any transactional email, since the email is judged as part of the flow.
• On a schedule for the flows that matter most. A saved review is a baseline, and the next run reports the delta.

Frequently asked questions

What is a flow walkthrough?

A flow walkthrough is an AI stepping through one specific journey on your site, the one you name, and reporting on every step: what broke, what confused, and what could be better. Unlike a full usability pass, which decides for itself which jobs a new user would try, a flow walkthrough follows the path you care about right now, so it's the right study when you've just changed checkout, or onboarding, or the invite flow, and want it checked end to end.

How do I describe the flow?

In a sentence, however you'd say it to a colleague: 'sign up, create a project, then invite a teammate.' The AI restates it as numbered steps before it starts, tells you what it will and won't do, and then walks it. You don't write a script or a test case.

Does it handle the emails the flow sends?

Yes, and that's the leg most reviews never reach. Cobalt gives the AI a disposable inbox, so a verification code, a welcome email, or an invite email arrives somewhere it can read. Each one is rendered as HTML and judged as a screen: does it arrive, does it look right, does the link work, does the copy match the page that sent it.

Does it need a browser?

Yes. Walking a flow means clicking through your product, usually behind a login, so this study runs from a coding agent that drives a real browser: Claude Code, Cursor, or Codex. From Claude chat it can cover the public, logged-out screens and will say so before it starts.

Is it safe to run on production?

It creates a throwaway account with its own inbox, never enters a password you didn't set, stops before any payment or irreversible action and reports the handoff as a finding, and never solves a human-verification challenge. It tells you the steps it's about to take before it touches the browser.

What do I get back?

A review at a shareable link with the verdict at the top, then the flow step by step: a screenshot per step, what happened, and a finding where something broke or confused, ranked by whether it blocked the goal, slowed it, or was merely cosmetic. Emails appear in the sequence at the point they were sent.