| name | visual-compare |
| description | Visually compare design options (colors, tokens, variants) in a live preview |
| argument-hint | ["what to compare"] |
| allowed-tools | Task |
Visual Compare
Two-phase workflow. The main thread handles user interaction.
Phase 1: Generate Options (subprocess)
Use the Task tool to spawn the comparison agent:
- subagent_type:
visual-compare-agent
- max_turns:
20
- prompt:
$ARGUMENTS\n\nRemember: complete Step 3 (Sampler Check) before generating options.
The agent will check the sampler registry (generating a new sampler if the target token isn't covered), then generate 4 options, write them to the preview page (data.json), and return a structured summary with the option labels and descriptions.
Phase 2: User Picks (main thread)
After the subprocess returns:
- Tell the user to check http://localhost:3000/dev/preview to compare visually
- Present the option labels from the subprocess summary and ask the user to pick
- If user picks "Other", re-run Phase 1 with adjusted prompt
Phase 3: Apply Choice (resume subprocess)
Resume the same agent using the Task tool's resume parameter:
- resume: the agent ID from Phase 1
- prompt:
User chose: <selected option label>. Apply the value to the target CSS file, clean up data.json, and return a final summary.
Report the final summary back to the user.
Autonomous Mode (programmatic use)
When another agent invokes visual-compare-agent with autonomous: true, the entire flow collapses into a single Task invocation — no user interaction or resume needed.
Task tool call:
subagent_type: visual-compare-agent
max_turns: 20
prompt: |
Compare CSS variable options:
- variable: --destructive
- file: apps/web/app/globals.css
- mode: dark
- context: error states and destructive action buttons
- autonomous: true
Remember: complete Step 3 (Sampler Check) before generating options.
The agent generates options, picks the best fit, applies it, cleans up data.json, and returns a SAMPLER: / CHOSEN: / APPLIED: summary. See the agent definition for full details on selection criteria and return format.