You walked a broken screen, dropped pins on the three things that are wrong, and typed a comment on each. Now you want Claude Code to fix them, one at a time, without you retyping any of it. The path is short: publish the review, grab its markdown, and paste it in with a clear instruction. Here is exactly how.
Capture the items before you publish
Each thing you want fixed should be its own item. If the header overlaps the logo on mobile, and the submit button is disabled when it should not be, those are two separate captures with two separate comments, not one paragraph covering both. Claude Code reads structure. When your review already reads as a list, the task list falls out of it.
Add a numbered pin where the problem sits on the screenshot. Pin 1 on the overlapping header, pin 2 on the dead button. Then type or dictate what is wrong and what you expect instead. "Header logo is clipped below 375px wide, should stay fully visible" beats "header looks off." If you are new to the capture flow, the step-by-step capture to markdown walkthrough covers the mechanics.
Publish and copy the markdown URL
Click Publish. The review saves and gets a short public address of the form /r/<slug>. That page is the human-readable version, fine for a teammate to read but not the one you hand an agent.
Add /markdown to the end. So if your review lives at /r/checkout-bugs, the markdown is at /r/checkout-bugs/markdown. Open it. You will see plain text: a heading per item, the comment you wrote, and references to the captured stills and their pins. No visual noise, no styling an agent has to parse around. This is the format a coding agent reads, and it is the whole reason to route feedback through a published link instead of pasting loose notes.
You do not need an account to publish. An anonymous review sticks around for 30 days, and if you sign in later you can claim it permanently. If most of your reviews feed an agent, signing in with a Google account or a magic link keeps them from expiring.
Hand the markdown to Claude Code
Open Claude Code in your project. Two ways to get the content in:
- Copy the rendered markdown from the
/markdownpage and paste it directly into the chat. - Give Claude Code the URL and ask it to fetch the page, if your setup allows it to read a link.
Then write the instruction that turns items into tasks. Something like: "Read the review below. Turn each item into a separate task. For each one, tell me the file you would change and the fix, then wait for me to approve before editing." Claude Code will parse the headings, treat each as a unit, and produce a numbered plan that maps back to your pins.
The reason this works cleanly is that your review already separated the problems. Item 1 becomes task 1. The pin coordinates and your "should be" language give Claude Code enough to locate the element and know when it is done. For the full setup, including how to keep the loop tight across several fixes, read the guide on getting feedback into Claude Code.
Review the plan before it starts editing
Do not let Claude Code fix all three items in one pass without a check. Ask it to lay out the plan first: which file, what change, per task. You will catch the case where it misreads a comment or targets the wrong component. This is where a specific comment pays off. If your item said "submit button stays disabled after all fields are filled," Claude Code knows to look at the validation logic, not the button styling.
Approve task by task. When it finishes one, check the result against your screenshot, then move to the next. If a fix is off, you already have the pin and the original comment to point back to. There is a list of common mistakes when handing UI feedback to Claude Code worth skimming before your first run.
When something new breaks, capture again
Fixed the header, but now the button sits too low? Capture the current state, publish a fresh review, and feed that markdown back. Each pass is a new link, so you keep a clean record of what was reported when. This is the loop that makes iterative building actually ship: capture, publish, hand off, verify, repeat.
The same markdown you gave Claude Code also exports as a PDF or a Word document, so if a client or a PM needs the human version, they get one from the same review. One capture, two audiences. If you work with other tools, the Claude Code page and the broader coding agents hub cover the same flow for whichever one you use.
Start your next review at the capture screen, pin the problems, publish, and add /markdown to the URL. That single string is what turns a walked screen into a task list Claude Code can work through.