Give specific instructions
The agent performs best with clear, scoped requests. Instead of “make the page look better,” try:
- “Change the header background to blue and increase the font size to 24px”
- “Add a loading spinner to the submit button while the form is submitting”
- “Refactor the
UserCard component to use Tailwind instead of inline styles”
Use DOM context selection
Click on elements in the browser viewport to attach them as context. This gives the agent the exact HTML, CSS, and component information it needs to make targeted changes.
You can select multiple elements before sending a message. The agent sees all of them.
Leverage workspace context
The agent is more effective when it understands your project:
- Generate
WORKSPACE.md — Go to Settings → Skills & Context files and click Generate for your workspace. This creates an auto-generated project analysis.
- Write custom skills — Add
.stagewise/skills/ folders with SKILL.md files to teach the agent project-specific patterns.
- Use
@ mentions — Reference files and tabs directly in your chat messages.
Review changes carefully
The agent writes code changes to disk immediately, but all edits are tracked in the diff system:
- Accept individual hunks you want to keep
- Reject hunks to revert them to the baseline
- Undo/redo entire tool calls if needed
- Open the full Diff Review page for complex multi-file changes
Keep conversations focused
Start a new chat for each distinct task. Long, multi-topic conversations can reduce agent accuracy. Use the New Chat shortcut to start fresh.
