AI and feedback
Why a client's comment makes a bad prompt, and what fixes it
A list of comments is not a prompt, and the gap between the two is where a round goes wrong. This page is about the artefact itself: what a prompt has to carry before an agent can act on it, which of those things plain text can never contain, and why the difference decides whether the agent edits the right button.
No credit card. Your client doesn’t need an account.
01
"Make it pop" is not a bad request. It is an unfinished one.
When a client writes "make it pop", they are looking at one specific element, on one specific page, at one specific screen width, with an intent that feels obvious to them. All four facts are in their head and none are in the message. Hand that message to an AI agent and it will still do something: pick an element, guess at pop, and produce a confident diff you throw away.
The fix is not a stricter briefing template your client will never fill in. It is capturing the missing facts at the moment the comment is written, then shaping the text into something an agent can execute without needing a follow-up question answered first.
02
What a prompt has that a list does not
A list of comments is a record of what people said. A prompt is an instruction to something that is about to act. Three things separate them, and none of the three are a matter of tone.
A target: every item names the thing to change, so the agent is not silently choosing between four buttons in the same section. An order: items are ranked and duplicates are counted once, so an agent working through a limited run spends it on the broken checkout rather than on a caption someone mentioned twice. A stopping rule: the prompt states what to do with an item that is ambiguous, which is to stop and ask rather than pick something plausible.
That last one is the piece people leave out, and it is the one that changes behaviour most. Without an explicit instruction to stop, a model has no safe way to fail. It will always prefer a confident answer to no answer, because that is what it was trained to produce. Give it permission to leave an item open and it will use it.
03
The same request, twice
As the client sends it
- "make it pop"
- No page, and you have three that could match
- No element, and there are four buttons in that section
- No screen width, so you cannot tell a layout bug from a taste call
- No priority, so the agent treats it like everything else in the list
- No way to resolve it without another message and another day
As a prompt an agent can run
- #7: increase the visual weight of the primary call-to-action in the hero
- Page: /pricing
- Element: .hero .cta-primary, pin at 48% / 60% inside it
- Desktop, 1440 by 900
- Screenshot: the exact view the client was looking at
- Grouped with the other two comments on the same element, counted once
04
The six things a usable prompt item carries
Take any one of them away and you can predict what breaks.
| Ingredient | What it looks like | What breaks without it |
|---|---|---|
| The element | A selector, plus where the pin sits inside it | The agent edits the wrong button |
| The page | The URL the client was actually on | The change lands in a template shared by five pages |
| The device | desktop, tablet or mobile, with the viewport size | A responsive complaint gets fixed at the wrong breakpoint, or never reproduced |
| The screenshot | An image of the pinned spot at the moment of the comment | You argue about what the page looked like last Tuesday |
| The intent | The request rewritten to be actionable, meaning untouched | The agent takes the words literally and enlarges text when the client wanted contrast |
| The priority | An order, with duplicates merged so one thing said twice counts once | The agent burns its run on a copy tweak and stops before the broken checkout |
05
What plain text cannot carry
This is the boundary, and no rewrite pass moves it: these things are missing from the words themselves.
- Which element each item refers to. Written feedback has no anchor, so a careful prompt tells the assistant to ask instead of guessing.
- Which page, unless your client happened to type a URL.
- Which device and viewport they were on when it looked wrong.
- What the screen actually looked like. No screenshot exists for a sentence someone typed into an email.
- Who said it, and whether two people reported the same thing from two different threads.
06
Where the other four ingredients come from
They are not something a smarter model recovers later. They only exist if they were recorded when the comment was made. That is the entire job of a review link: your client opens one URL, needs no account, clicks the element they want changed, and types. Page, element, pin position, viewport, device type and a screenshot get attached without them doing anything extra.
Same for the Chrome extension when the page sits behind a login, on a staging URL or on localhost, and for the embeddable widget when you want feedback to be leavable in place on a live site at any time. Three ways in, one prompt out.
And if all you ever get is three bullet points in an email, write the prompt by hand. There is no reason to move a two-comment project into a review tool when a paragraph does the job.
Before you count on it
Writhink is in open beta: everything described here is available and free, no credit card. Paid plans are announced but not purchasable yet. The one limit that applies today is five AI prompt generations a month per organization.
FAQ
Questions people actually ask
- What can I paste into it?
- Anything up to about 8,000 characters: an email thread, chat messages, notes from a call, a bulleted list. It handles the common case of a client writing one long unpunctuated paragraph by falling back to splitting on sentences.
- Which AI should I paste the result into?
- Whichever one has your code. The output is plain Markdown addressed to an assistant, so it behaves the same in ChatGPT, Claude, Cursor, Copilot, or the chat box of an AI builder like Lovable or v0.
- Does the prompt just tell the AI to do it?
- It tells the assistant to merge duplicates, apply the changes one page at a time, return a checklist ordered by priority, and stop rather than guess on anything ambiguous. That last instruction earns its keep: one item guessed wrong costs you two rounds with the client.
- How is this different from the prompt Writhink generates inside a project?
- Same instruction head, same treatment of unclear items, same source file in the codebase. The difference is context. Inside a project each item carries its element, page, device type and screenshot, and the numbers match the pins your client can see on screen.
- Do you keep what I paste?
- The instant version never leaves your browser. The optional AI rewrite sends the text to the model to be rewritten and returns the result; it is not attached to an account, because there is no account.
Start
Send one link. Get feedback you can build from.
Your clients comment right on the page — no account, no screenshots, no voice notes. Writhink sorts the comments and writes the prompt for your AI.