Making Cursor’s agent write plans as artifacts so I can comment on them inline

Showcase Built for Cursor
, ,
1

Long plans in Cursor are hard to give feedback on: to fix one part I had to copy the paragraph into chat, quote it, and write my note under it, over and over. What works for me now is asking the agent to create an artifact, meaning the plan saved as a separate document in the IDE, where I can select any text, comment right there, and send all comments back in one go. I built an extension for this, called AI Artifacts.

GitHub: [link] · Step-by-step write-up: [dev.to link]

Create an artifact
artifact-opens

Leave inline feedback
select-comment

Pick it up in a new chat
reconnect

Here is how I use it. I ask the agent to “create an artifact” for the plan. It opens in the editor as a document, and I select any paragraph, list item, code block or table cell and write a comment (Enter saves it). When I’ve gone through the whole plan I click Review, and every comment goes back to the agent at once. It answers questions in chat, changes the document only where I asked for changes, and starts the next round. If the plan looks right, Proceed approves it and the agent starts on it, and Just save keeps it without running anything.

I also use it for things other than plans, like asking the agent to write up a workaround after I finish a feature.

It works through MCP plus a small skill. Run “AI Artifacts: Install All Detected Integrations” from the command palette, reload the window and start a new chat. I’ve used it with Cursor, Codex and GitHub Copilot, and it also works with Claude Code and Cline.

A few things to know before you try it. This is not Cursor’s built-in Plan Mode: the agent creates a separate document only when you ask for an artifact. It only works in local desktop editors for now, so no Remote SSH, WSL, dev containers or Codespaces. Artifacts are stored in ~/.ai-artifacts/ and deleted after 30 days by default (configurable), so use Just save or Copy Markdown for anything you want to keep. After updating the extension you need to run the installer again and restart the agent. Everything stays local and there is no telemetry.

How do you give feedback on plans today? I’d like to hear what’s missing for your workflow.

I’m the author, and I built the project with AI assistance.

Inline Comments in Plan Mode (like in Google's Antigravity)
Plan Mode is inefficient
In-line Comments/Annotations in Plan Mode (a la Antigravity)
Planning & Execution Artefacts like Google Antigravity
Planning mode - inline comments
5

nice, the batch review round is the part plan mode is missing. one gap worth tackling next is the other end. after Proceed, diff what the agent actually changed against the approved plan, files touched that the plan never mentioned, steps skipped. approving the plan is only half the review if nobody checks the agent stuck to it. separately, a PRI (Production Reliability Index) score on what it shipped, which is what we work on, tells you how closely to read the result.

1 Like
6

Thanks, fair point. Right now it stops at Proceed, so nothing checks that the agent followed the plan. For now you can ask the agent to write a follow-up artifact comparing what it did against the approved plan. It shows up next to the plan in the artifact viewer: click the “Conversation Artifacts” button in the viewer’s header to see every artifact from the same chat.