agentation-self-driving

Agentation Self-Driving Mode

Autonomously critique a web page by adding design annotations via the Agentation toolbar — in a visible headed browser so the user can watch the agent work in real time, like watching a self-driving car navigate.

Launch — Always Headed

The browser MUST be visible. Never run headless. The user watches you scan, hover, click, and annotate.

Preflight: Verify agent-browser is available before anything else:

command -v agent-browser >/dev/null || { echo "ERROR: agent-browser not found. Install the agent-browser skill first."; exit 1; }

Launch: Try opening directly first. Only close an existing session if the open command fails with a stale session error — this avoids killing a browser someone else is using:

Then verify the Agentation toolbar is present and expand it:

"Block page interactions" must be checked (default: on).

eval quoting rule: Always use [class*=toggleContent] (no quotes around the attribute value) in eval strings. Do not use double-bang in eval because bash treats it as history expansion. Do not use backslash-escaped inner quotes either, as they break unpredictably across shells.

Critical: How to Create Annotations

Standard element clicks (click @ref) do NOT trigger annotation dialogs. The Agentation overlay intercepts pointer events at the coordinate level. Use coordinate-based mouse events — this also makes the interaction visible in the browser as the cursor moves across the page.

@ref compatibility: Only click, fill, type, hover, focus, check, select, drag support @ref syntax. The commands scrollintoview, get box, and eval do NOT — they expect CSS selectors. Use eval with querySelector for scrolling and position lookup.

If no dialog appears after clicking, the toolbar may have collapsed. Re-expand (only if collapsed) and retry:

Building CSS selectors from snapshots

The snapshot shows element roles, names, and refs. Map them to CSS selectors:

When in doubt, use a broader selector and verify with eval:

The Loop

Work top-to-bottom through the page. For each annotation:

1. Scroll to the target area via eval (scrollIntoView)
2. Pick a specific element — heading, paragraph, button, section container
3. Get its bounding box via eval (getBoundingClientRect)
4. Execute the coordinate-click sequence (mouse move → mouse down → mouse up)
5. Read the full snapshot output to find dialog refs at the bottom
6. Write the critique (fill @ref) and submit (click @ref)
7. Verify the annotation was added (see below)
8. Move to the next area

Verifying annotations

After submitting each annotation, confirm the count increased:

If the count didn't increase, the submission failed silently — re-snapshot and check if the dialog is still open.

Aim for 5-8 annotations per page unless told otherwise.

What to Critique

Critique Style

2-3 sentences max per annotation:

• Specific and actionable: "Stack the install command below the subheading at 16px" not "fix the layout"
• 1-2 concrete alternatives: Reference CSS values, layout patterns, or design systems
• Name the principle: Visual hierarchy, Gestalt grouping, whitespace, emphasis, conversion design
• Reference comparable products: "Like how Stripe/Linear/Vercel handles this"

Bad: "This section needs work" Good: "This bullet list reads like docs, not a showcase. Use a 3-column card grid with icons — similar to Stripe's guidelines pattern. Creates visual rhythm and scannability."

Install

The skill must be symlinked into ~/.claude/skills/ for Claude Code to discover it:

Restart Claude Code after installing. Verify with /agentation-self-driving — if it loads the skill instructions, the symlink is working.

Troubleshooting

• "Browser not launched. Call launch first.": Stale session from a previous run — run agent-browser close 2>/dev/null then retry the --headed open command
• Toolbar not found on page: Agentation isn't installed — run /agentation to set it up first
• No dialog after clicking: Toolbar collapsed — re-expand with the state-aware eval (check [class*=expanded] first), retry
• Wrong element targeted: Click Cancel, scroll to intended element, retry with correct coordinates
• Add button stays disabled: Text wasn't filled — re-snapshot and fill the textbox
• Page navigated: "Block page interactions" is off — enable via toolbar settings
• Annotation count didn't increase: Submission failed — dialog may still be open, re-snapshot and check
• Interrupted mid-run (Ctrl+C): The browser stays open with whatever state it was in. Run agent-browser close to clean up before starting a new session

agent-browser Pitfalls

These will silently break the workflow if you're not aware of them:

Rule of thumb: @ref works for interaction commands (click, fill, type, hover). For everything else (eval, get, scrollintoview), use CSS selectors via querySelector in an eval.

Two-Session Workflow (Full Self-Driving)

With MCP connected (toolbar shows "MCP Connected"), annotations auto-send to any listening agent. This enables:

• Session 1 (this skill): Watches the page, adds critique annotations in the visible browser
• Session 2: Runs agentation_watch_annotations in a loop, receives annotations, edits code to address each one

The user watches Session 1 drive through the page in the browser while Session 2 fixes issues in the codebase — fully autonomous design review and implementation.