Product feedback for Replit Agent

This is the playbook for getting visual product feedback into a Replit Agent session without losing context across iterations. For the same workflow applied to other tools, see the full feedback for AI coding agents hub.

The problem with feedback in Replit Agent workflows

Replit Agent lives inside the Repl. You watch the preview pane on the right, you chat with the agent on the left, and the agent rewrites files in your project as you go. The usual feedback path is one of two bad options. You paste a screenshot into the chat and type a sentence under it. Or you write a paragraph describing what's wrong without a screenshot at all.

Both lose context the moment the next iteration starts. Pasted screenshots scroll out of the chat history. Typed descriptions don't reference selectors or components. When the agent rebuilds a section, the feedback that prompted the rebuild is already half-forgotten. You end up retyping the same critique three iterations in a row, each time slightly different.

What Replit Agent actually needs is a structured artifact inside the project that the agent can reread on demand.

The CobaltCapture workflow with Replit Agent

Open your Repl and put the preview in its own browser tab so you can capture it cleanly. Open cobaltcapture.com in another tab and hit Capture screen. Pick the preview tab. Drag a box around the broken part of the UI, then hit Dictate and talk through what's wrong out loud. "The nav collapses one breakpoint too early, the hamburger should kick in at 768px not 900px." Repeat for each finding. Hit Publish.

You get a public review URL and a one-click markdown export. Now drop the file straight into the Repl filesystem. From the Repl shell:

curl -fsSL https://cobaltcapture.com/r/<slug>/markdown -o feedback.md

feedback.md now lives next to your index.html and package.json. Open the Replit Agent chat and reference it:

Read feedback.md and address items 1, 2, and 3 in order. For each item,
summarize what you understood, then apply the fix. Pause for confirmation
before moving to the next item.

Replit Agent reads the file as a first-class project artifact, follows the embedded image URLs for visual context, and works through the findings in order.

Example prompt

I just reviewed the preview and the issues are documented in feedback.md
at the project root. Please address each item in order, show me the diff
before applying, and run the preview to confirm.

Why this works for Replit Agent

Replit Agent operates inside a single, self-contained Repl. The project filesystem is the source of truth, and the agent already treats files in the project as canonical context. A feedback.md at the root is exactly that kind of file, discoverable, persistent, and rereadable across every iteration in the session.

Survival across sessions matters here more than it does in most editor-based agents. Replit sessions get reopened a day later when you come back to keep iterating, and pasted chat context is gone by then. The file isn't.

Dictation matters too. Replit Agent's prompts are short by design, the chat input rewards brevity. Dictating the nuance into the review document means the prompt itself stays short ("read feedback.md, fix items 1-3") while the agent still gets the long-form context it needs to produce a real fix instead of a shallow one.

Alternatives and tradeoffs

You can paste screenshots straight into the Replit Agent chat. It works for one or two findings. Past that, the images scroll out of view, your descriptions lose their anchor, and the record vanishes when the session ends.

You can type the description into chat without a screenshot. That's faster than taking a screenshot but slower than dictating, and it skips the visual context the agent needs to identify the right component.

You can file the findings into an external ticket tracker and reference them in chat. That's overkill for the fast iteration loop Replit Agent is built for, you're switching tools and writing tickets for something you're about to fix in the next two minutes.

The point of CobaltCapture is that the artifact lives in the project, gets read by the agent the same way every time, and is structured the same regardless of who reviewed. For the broader pattern this sits in, see agent-readable feedback as a workflow.