You finished a design review in Cobalt. Six items, screenshots and all, published at a short link. Now you want it as a ranked to-do list you can actually work from. You do not need to copy anything out or re-type a word. Paste the link into ChatGPT and ask.
Here is the full example, start to finish, with the real link, the real prompt, and the real shape of what comes back.
The review we start with
Say you ran a pass on a checkout page and published it. The link is https://cobaltcapture.com/r/checkout-pass-4k2. Six items sit under it:
- Item 1: Pin on the "Place order" button, comment "disabled state looks identical to enabled, I tried clicking it three times."
- Item 2: Pin on the promo field, "invalid code shows no error, the field just clears."
- Item 3: Screenshot of the phone view, "total is cut off below the fold, you have to scroll to see what you pay."
- Item 4: "Shipping estimate says '3-5 days' but the confirmation email says '1 week'."
- Item 5: "Card logos are low-res on retina."
- Item 6: Free-floating note, "consider autofilling the country from the IP."
Each item carries its screenshot, any pins, and the comment. That is what Cobalt publishes, and it is also what ChatGPT gets to read.
Why ChatGPT can read the link at all
Every published review has a plain-text version at /r/<slug>/markdown. When an assistant opens a /r/ link, Cobalt serves it that markdown automatically, screenshots included. So you do not paste the markdown URL and you do not export anything. You paste the normal link and ChatGPT receives a structured document: a heading per item, the comment text, and the image. The mechanics of that format are covered in markdown screenshots and the terms you need, but you do not have to think about any of it to use it here.
This is the difference between a screen recording and a captured review. A video forces the reader to watch. A review link hands the reader text it can parse. If you want the longer argument, see why a Loom video does not work as agent input.
The prompt
Open ChatGPT and paste this:
Read this review and turn it into a ranked task list. Rank by user impact, highest first. For each task give me a one-line title, the item it came from, and whether it is a bug or an improvement. https://cobaltcapture.com/r/checkout-pass-4k2
No prompt engineering, no system message, no uploaded files. The ChatGPT workflow for reading a review link is just this: give it the link and tell it what shape you want back.
What ChatGPT gives back
Because it read the comments and saw the screenshots, it ranks on what each problem costs a user, not on item order. A typical result:
- Fix the disabled "Place order" state so it is visibly different (bug, item 1). A user who cannot tell the button is disabled cannot buy.
- Show an error when a promo code is invalid instead of clearing the field (bug, item 2). Silent failure reads as the page being broken.
- Reconcile shipping estimate between page and confirmation email (bug, item 4). Contradictory delivery promises undercut trust at the worst moment.
- Keep the order total visible above the fold on phone (bug, item 3). People want to see what they pay before they commit.
- Serve retina card logos (improvement, item 5). Cosmetic, low effort.
- Autofill country from IP (improvement, item 6). Nice-to-have, noted as a suggestion.
Notice it caught that item 6 was framed as a suggestion ("consider") and ranked it accordingly, and that item 4 is a cross-surface mismatch rather than a layout bug. That reading comes from the comment text, which is why a clear one-line comment under each capture is worth the few seconds it takes. The habit is worth building: see the one comment to write under a quick shot.
What to do with the list
You now have a prioritized backlog in the chat. Common next moves:
- Ask ChatGPT to rewrite the top three as GitHub issue bodies. It already has the screenshots and the pin context.
- Ask it to group the bugs separately from the improvements so you can ship fixes first.
- Paste the list into your tracker as-is. The item references point back to the review, so anyone can open the link and see the exact pin.
The ranking is only as good as the review. If a comment is vague ("this looks off"), ChatGPT ranks it low or guesses, because there is nothing to rank on. Pin the exact spot and say what is wrong in a sentence. The same principle holds whether a person or an assistant reads it, and it is the core of writing feedback specific enough to act on.
Starting from nothing
If you do not have a review yet, start a new review in a browser tab. There is no install, no extension, and no password. Capture the screen, pin the spots, type or dictate your notes, and publish. You get unlimited reviews free to try; Cobalt asks for an email only when a review is ready so it can send the link. After 7 days a review goes read-only for you (the shared link keeps working), and Pro at $5 a month keeps every review editable with no cutoff. Full terms are on the pricing page.
One link in, a ranked task list out. That is the whole loop.