DEV Community

Cover image for This extension lets your AI agent create artifacts for plans, workarounds and designs in Cursor, Codex and GitHub Copilot
Huy Nguyen
Huy Nguyen

Posted on Edited on

This extension lets your AI agent create artifacts for plans, workarounds and designs in Cursor, Codex and GitHub Copilot

This guide uses an extension called AI Artifacts to let Cursor, Claude Code, Codex, Windsurf, GitHub Copilot, Cline, or Antigravity produce a plan, workaround note, or design as a separate document—an artifact—inside your IDE. Instead of scrolling through a narrow chat sidebar or reading raw Markdown syntax, you read a rendered preview (with formatted tables, syntax-highlighted code, and Mermaid diagrams), comment on exact passages, and send all feedback back in a single step.

What you will need

  • VS Code 1.95 or newer, or a compatible editor such as Cursor, Windsurf, Antigravity or VSCodium
  • An AI agent that supports MCP and skills (Cursor, Claude Code, Codex, Windsurf, GitHub Copilot, Cline or Antigravity)
  • Node.js available as node in your PATH

Step 1: Install and connect the extension

Install AI Artifacts from the VS Code Marketplace or Open VSX. The extension communicates with your agent through the Model Context Protocol (MCP) and a small skill that defines how documents are created and updated.

  1. On first activation, click Run Installer in the prompt—or open the Command Palette (Ctrl+Shift+P / Cmd+Shift+P) and run AI Artifacts: Installer. This preselects detected clients on the machine and leaves unrelated MCP settings and skills unchanged; press Enter (or click OK) to configure them.
  2. Click Reload Window when prompted.
  3. Restart your AI chat extension or agent, then start a new chat. A chat that was already open cannot load tools installed after it started.
  4. Run AI Artifacts: Verify All Integrations. The skill and the base runtime should both report Ready.

_Running AI Artifacts: Installer_

Step 2: Ask the agent to create an artifact

The agent creates a review document only when it is asked to. A request for "a plan" on its own produces a plan in the chat. Include the word artifact in the prompt:

Create a dummy plan with: a mermaid diagram, a comparison table, code blocks,
a task list, and open questions. Create an artifact for me to review.
Enter fullscreen mode Exit fullscreen mode

You can also name the skill directly: create-review-artifact (in Codex, $create-review-artifact).

_The agent creates the artifact and the review editor opens_

The review editor opens in your focused IDE window automatically, and the agent also shares a clickable link to the artifact in chat so you can open it directly.

Step 3: Review the document with inline comments

Read the plan in the rendered preview—with formatted tables, code blocks, and Mermaid diagrams instead of raw Markdown symbols. Select any paragraph, heading, list item, quote, code block or table cell, type a comment in the popover, and press Enter to save it (Shift+Enter inserts a new line). Saved comments are listed in the Comments drawer, and the Review button shows how many are pending.

In the example plan, four comments were left:

  • Interface sketch: add more detail.
  • Tasks: add more detail.
  • Acceptance criteria: delete the section, which is unnecessary for a feature this small.
  • Rollback: delete the section for the same reason.

_Selecting text and leaving a comment_

Comments can also be questions, such as asking why a particular approach was chosen.

Step 4: Submit the review

Four decisions are available in the editor header:

  • Review sends all saved comments to the agent and starts the next round. It requires at least one comment.
  • Proceed approves the document. For a plan, the agent begins carrying it out immediately.
  • Just save ends the review without executing anything. It suits cases where you only want to keep an idea: the agent saves the document to a file so you do not have to copy it manually.
  • Copy Markdown copies the raw text to the clipboard without changing anything.

The example plan still needed changes, so Review is the correct choice. The agent acknowledges the comments in the chat, summarizes what it is applying, and opens round 2.

Step 5: Verify round 2

The agent answers questions in the chat and revises the document only where changes were requested. If a review contains only questions, the document stays byte-for-byte identical and the round still advances.

Round 1 of the example, with four comments saved:

_Round 1: four comments saved. Tasks, Acceptance criteria and Rollback are highlighted; the Interface sketch comment is just above_

Round 2, after the agent applied them:

_Round 2: a longer, more detailed task list, and no acceptance criteria or rollback sections_

The task list grew from 7 items to 11 with more concrete steps, the API example gained its response, and the two sections flagged for deletion were removed. The review can continue for as many rounds as needed, ending with Proceed or Just save.

Step 6: Resume a review

An artifact remains available after its editor tab is closed or the chat ends, so a review can be resumed without starting over.

In the same chat, ask the agent to reopen the artifact.

In a new chat:

  1. Run AI Artifacts: Search Artifact from the Command Palette and type part of the title.
  2. Select the artifact to open it in the review editor.
  3. Click Inspect artifact in the editor header. This copies a prompt that points to that exact artifact.
  4. Paste the prompt into the new chat.

The agent reads the saved state and reports the current review round. Reconnecting never repeats an action that was already approved.

_Search an artifact, copy the Inspect prompt, paste it into a new chat, and the agent reads the saved state_

You can also browse and reopen recent documents directly from the AI Artifacts sidebar in the Activity Bar under Recent Artifacts.

_Reopening an artifact from Recent Artifacts in the AI Artifacts sidebar_

Step 7: Have a second AI review your plan

Because Inspect artifact lets any supported AI agent read the exact same artifact, you can get a second opinion on an important plan before you click Proceed:

  1. While viewing the plan drafted by Agent 1 (for example, in Cursor), click Inspect artifact in the editor header.
  2. Open a chat with a second AI agent (for example, Claude Code or Codex), paste the prompt, and ask it to review the proposal:
   [Pasted Inspect artifact prompt]
   Review this plan for edge cases, missing steps, or simpler alternatives, and write your feedback into the artifact.
Enter fullscreen mode Exit fullscreen mode
  1. Agent 2 reads the artifact and writes its review comments directly onto the document (or replies in chat), so you and Agent 1 can refine the plan in the next round before executing code.

Step 8: Use artifacts for other document types as well

Artifacts are not limited to plans. You can use the same review workflow for requirements, designs, specs, analysis reports, or workaround notes:

Create a requirements artifact for the webhook retry policy, including edge cases
and acceptance criteria, so I can review it.
Enter fullscreen mode Exit fullscreen mode

You can also ask the agent to create a design, spec, analysis, report, or workaround artifact whenever you want to review structured documents outside the chat thread.

What's next?

Reviewing a single artifact is all you need when a task revolves around one document. When a session produces multiple artifacts that you want to browse in one place—or when you want to carry shared instructions and context into fresh chats—move up to Work Sessions:

Links

Feedback is welcome in the comments or as an issue on GitHub. Please include the editor and agent you are using.

Top comments (0)