FIX WHAT IS BROKEN
Fix a multi-step flow that breaks midway
In a multi-step flow the step that visibly fails is often not the step that is wrong. A value is dropped on step two and nothing goes visibly wrong until step five. If you only record step five, you have handed the agent the symptom and thrown away the cause.
Quick answer
To fix a multi-step flow that breaks midway, record one continuous run from the true start of the flow with test data, saying values out loud as you enter them. Paste the Clipy link into your AI agent, such as Claude Code, Codex or Cursor, and ask for the earliest step where the state stops matching what the final step needed. The fix goes there.
Product walkthrough. Tap for sound.
What to record
Start at the true beginning of the flow, even if the first steps are boring.
Go through it the way a real user would, including any values you type.
Let each step's result be visible before you move on.
When it fails, stop and say what you expected the flow to have carried forward.
Say this while you record
Say the values out loud as you enter them. That puts them in the transcript, so the agent can see what you entered and what came out the other end.
The prompt to paste
Replace the link with your own recording's share URL, then paste this into Claude Code, Cursor, Codex, or whichever agent you connected.
Read https://clipy.online/video/<id> It is a recording of a multi-step flow that fails at the end. List the steps as you see them, with timestamps, and what state each one appeared to leave behind. Then tell me the earliest step where the recording shows something that does not match what the final step needed. Find that in the repo and propose the fix there, not at the point where the error surfaced.
Pointing the agent at the earliest divergence rather than the visible failure is the whole trick with flows.
What the agent gets from that link
A summary of what happened
A short written account of the recording, so the agent starts with the shape of the request rather than a wall of raw text.
A timestamped transcript
Your narration with times against it, when audio was captured and processed. This is where your reasoning lives, and it is usually the part that decides whether the agent fixes the right thing.
Key moments with frames
Selected instants from the recording with an image of the screen at that point, so the agent can look at the state you were describing instead of guessing at it.
Pointer and click evidence
Where you clicked and what you were pointing at, with the control's accessibility label when the capture recorded one. It turns "this button" into a specific element.
What changes
| Writing it down | Recording it instead | |
|---|---|---|
| The bug report | "Checkout fails at payment." | The whole run, including what was entered at step two. |
| Where the agent starts looking | At the payment step, which is fine and often wrong. | At the earliest point where the recording shows a mismatch. |
| Round trips before a fix | Several, mostly spent asking what you typed. | Usually one. |
Tips
Connect your agent once
Run the setup at /agent-setup once. After that every Clipy share link works in the same session you are already in.
Do not type real credentials or card numbers
The recording captures what is on screen. Use test data. If something sensitive is briefly visible, do not share the link and record it again.
Pick the window you mean
On a Mac, window capture records only the window you pick, even when other windows cover it. The Mac's window-and-screen-area mode, which includes the camera bubble and Flare, and the Windows app record the window's area of the screen, so close anything you do not want in that area before you start.
Long flows still need one continuous recording
Two recordings of two halves lose the handover, which is usually where the bug is. Keep it in one capture even if it runs a few minutes.
Common questions
What if the flow spans two apps or a redirect?
Record it as one capture across both, using display or area capture rather than a single window. The recording just follows what is on the screen.
What about steps that need an email confirmation?
Record those too, on screen, with a test inbox. The waiting time is uninteresting but harmless, and the agent gets the full chain.
Should I record a working run as well?
If you have one, yes, it is a strong second link. "Here is it working, here is it failing" is the fastest way to point at the difference.
Other jobs
The same job, elsewhere on Clipy
This page is the job. These cover the same ground for a specific agent, a specific team, or as a tool you can open right now.
The full list is on the use-case index. If your agent is not connected yet, start at agent setup.