Tell AI Exactly What to Fix. Lasso Is the Handoff Layer for AI Agents
Parry Gill · Oct 31, 2026 · 6 min read
Here is a scene most product teams will recognize. A designer spends twenty minutes writing a Slack message to explain one spacing bug. Two screenshots, a red box drawn in Preview, three paragraphs of explanation, and then the whole thing gets pasted into Cursor. Cursor fixes the wrong div. Confidently, and with a very nice commit message.
Nothing about that was the agent's fault. It did exactly what the message said. The message just never said which element, on which page, in which state.
That gap is what this post is about. The agents got fast. The way we hand them work did not.
Building got fast. Explaining did not.
Cursor, Claude Code, Codex, Lovable and v0 will happily rewrite a component or rebuild a pricing page before your coffee cools. Where things slow down now is the bit in between, where a person who noticed something has to describe it to a model that cannot see their screen.
A human developer can lean over and ask "which button?" An agent is more like a brilliant intern who has never once asked a clarifying question. It guesses, and it guesses with total confidence.
The fixes that land on the first try tend to have four things in the feedback behind them. The page and the exact element. A picture of what it looked like. What the person actually wants. And whether the item is still open or was settled in a thread somewhere. Screenshots in Slack give you one of those, maybe two on a good day.
What we mean by a handoff layer
We call Lasso a handoff layer because that is the job it does. People point at the real page and say what they want. Lasso keeps all of it in one place while the team argues it out. Then it hands the result to an AI in a form the AI can use.
Pointing at the real page
You install the Chrome extension, open any page and click the thing that is wrong. It works on production, on staging, on a client's site, even on a dashboard behind your login, because it runs in your browser and needs nothing installed on the site.
When you click, Lasso quietly saves the page address, a selector for the element, the text inside it and a screenshot of that exact spot. You just type the comment.
That last part matters more than it sounds. A lot of AI work starts on pages you don't control. "Match how this competitor does onboarding." "The vendor widget is overlapping our footer." Tools that need a script on your own site can't help there at all.
Letting the team settle it first
Comments go into a collection that you share with whoever needs to see them. People reply, tag each other and mark things resolved. By the time anything reaches an agent, the duplicates are gone and the bad ideas lost the argument.
Skip this step and you will meet my least favorite AI bug, where the agent carefully implements a request that someone took back two replies later.
Handing it over
When a collection is ready, Export for AI turns it into a brief. You can copy it as Markdown for Cursor, Claude Code or ChatGPT. You can download a PDF with every screenshot inside, which is handy for chat tools that read images but can't open links. Or you can switch on a private link that an agent fetches directly, and switch it off again whenever you like.
The Markdown starts with a short note telling the model how to read the file, then lists each comment like this.
### #1 | maya | 2026-10-30 14:12 UTC | OPEN
Page: https://staging.acme.com/pricing
Element: section.plans > div.plan-card:nth-child(2) > a.btn
Element text: Start free trial
Comment: This CTA should say "Start your 14-day trial" and use the primary style.
Screenshot: https://app.lasso.team/ai/…/img/…
Replies:
- dev (2026-10-30 15:02 UTC): Primary style is the orange gradient one in the hero.
### #2 | sam | 2026-10-30 14:20 UTC | RESOLVED
Page: https://staging.acme.com/pricing
Element: h1.page-title
Comment: Headline wraps badly on mobile.
Every item carries where, what it looked like, what's wanted and whether it's still open. Nobody has to write that up by hand.
Things an agent can't see for itself
People ask why they can't just paste a URL into the agent. Sometimes that works. Often it doesn't, because a model fetching a page gets the raw HTML rather than what you saw.
Charts drawn on a canvas have no text to read. Tables that load after the page renders are often missing. Anything behind a login is invisible. Hover states, open menus, error messages and the mobile layout never show up at all.
Lasso records what the person was looking at, in their own signed-in browser, at the moment they clicked. That is context the agent's tools can't go and fetch on their own.
A loop that works
The loop we recommend is pretty boring, which is the point.
- Someone clicks through the build and pins comments wherever something looks off.
- Someone else tidies the collection and settles open questions in the replies.
- We export the collection and paste it into the agent with a short instruction.
- We reload the page with Lasso open, check each pin, resolve what's done and export again.
The instruction doesn't need to be clever.
Work through every OPEN item in this Lasso export. Use the Element and Element text lines to find each element. Treat RESOLVED items as context only, follow the latest reply where it changes the request, and tell me anything you couldn't find.
Because the pins stay attached to the live elements, the second round is as precise as the first.
Where it pays off
Reviewing prototypes from Lovable and v0 is a natural fit, since the last twenty percent always needs a human eye. Design reviews are another, where a designer pins spacing and copy issues straight onto the build instead of leaving a Figma thread. For QA, every bug arrives with the element and a screenshot already attached. For marketing, a dozen copy changes across a dozen pages become one export.
It is also a decent way to brief an AI on a competitor. Annotate their pricing page, export the collection, and ask for a first draft of yours.
What changes when you work this way
The most obvious change is that agents get it right more often on the first pass, because they're told the page, the element and what it looked like rather than left to guess. You spend less time writing long prompts that describe a screen, and almost no time on the "no, the other one" back and forth.
Reviews also leave a trail. Every request, the discussion around it and whether it was done stays attached to the page, so the next person (or the next agent) can see the history.
People decide, the agent does the typing
AI is very good at doing the work. It is much worse at knowing what good looks like for your product and your customers. That part is still a person looking at the real page and saying "this one, here, more like that".
Lasso is our attempt at making that moment quick and precise, so nothing gets lost on the way to the agent. The export only contains the collection or site you pick. The private link is unguessable and stops working the second you turn it off. Resolved items are clearly marked, so they get read as background instead of being done twice.
Try it
Export for AI is part of Lasso's free plan, along with unlimited comments, replies and mentions. Pin a few comments on something you're building this week, hand the export to your agent and see how often it gets the right element the first time.