| name | ai-surface-smoke-research |
| description | Run the shared browser, desktop, and optional voice-agent smoke workflow across ChatGPT, Claude, Reddit, X, and other requested AI surfaces. Use for voice-triggered or typed end-to-end tool tests, AI chatter collection, browser scrolling, desktop focus, visible-action proof, and concise tool-grounded summaries. |
AI Surface Browser and Desktop Smoke Test
Run a quick end-to-end Prometheus test that proves desktop focus, browser navigation, social/search collection, and summary generation all work in one flow.
When to Use This
Use this when Raul wants a lightweight test workflow, not a deep research project. The point is to exercise the system: focus native apps, use the browser, collect current AI chatter from Reddit and/or X, and summarize what was found.
Good topics include Claude, OpenClaw, Hermes, Codex, AI agents, or any AI tool names Raul gives.
Default Workflow
-
Focus desktop AI apps
- Find/focus ChatGPT with
desktop_find_window({ name:"ChatGPT" }) then desktop_focus_window({ name:"ChatGPT" }).
- Find/focus Claude with
desktop_find_window({ name:"Claude" }) then prefer exact returned handle when needed, because Chrome/X titles may include the word Claude.
- If a name collision happens, use
desktop_find_window results and desktop_window_screenshot({ handle }) to verify the real Claude app.
-
Exercise the browser
- Open the user-requested search or AI surface. For a lightweight execution-only test, open X live search and perform one ordinary scroll.
- For a collection/research test, use
browser_scroll_collect rather than manual scroll loops.
-
Search Reddit when research is requested
- Use browser automation, not plain web search, when the test goal is browser execution.
- Default URL:
https://www.reddit.com/search/?q=<encoded query>&type=link
- Use
browser_scroll_collect for 3-5 scrolls.
- Capture post titles, subreddits, rough votes/comments, and repeated themes.
-
Search X when research is requested
- Default URL:
https://x.com/search?q=<encoded query>&f=live
- Use
browser_scroll_collect for 3-5 scrolls.
- Capture authors, handles, timestamps, text, links, and visible metrics when available.
- Do not like, reply, repost, bookmark, follow, or post.
-
Fallback web research only when browser/social collection is blocked
- If the point of the run is browser execution, do not replace the browser steps preemptively.
- If Reddit/X cannot be collected because browser or auth is blocked, use
web_search({ fetch_top_k: 2-4 }) for Reddit/web mentions or web_fetch_batch for selected URLs.
- Clearly label this as fallback research, separate from the browser/desktop smoke-test result.
-
Summarize or prove completion
- Report after the actual tool sequence completes.
- Keep it concise unless Raul asks for depth.
- For execution-only or voice-triggered tests, keep the reply to one short completion sentence and send fresh screenshot proof to the origin surface when requested.
- For research tests, include surfaces tested, app focus result, source highlights, and a short read on the AI chatter.
Default Query
If Raul does not provide a query, use:
Claude OpenClaw Hermes AI
Output Template
Done — focused ChatGPT and Claude, searched Reddit + X for `<query>`, and collected live results.
Quick read:
- Reddit: <2-4 bullets about repeated themes>
- X: <2-4 bullets about repeated themes>
- Overall: <one-sentence synthesis>
Tool check: browser collection worked; desktop focus worked/partially worked with <note if any blocker>.
Rules
- This is a test flow, so skip
declare_plan unless Raul explicitly asks for a plan.
- Do not take external social actions. Read-only only.
- Do not read or alter ChatGPT/Claude chat content unless Raul explicitly asks.
- Prefer exact window handles when desktop window names collide with browser page titles.
- If Reddit is logged out, continue; logged-out search results are still usable.
- If X auth/search fails, report the exact blocker and continue with Reddit or web fallback.
- Use fresh visual evidence for ambiguous desktop focus.
Non-Triggers
Do not use this for:
- Deep competitive intelligence reports.
- Posting, liking, replying, or outreach.
- ChatGPT coding handoffs.
- Voice-model latency, interruption, or audio-quality benchmarking that does not exercise browser/desktop tools.
Quality Check
A successful run has evidence that:
- ChatGPT was focused or explicitly found missing.
- Claude was focused or explicitly found missing.
- Reddit and/or X was opened in browser automation.
- At least one collection pass returned real content.
- Final summary is grounded in collected text, not vibes.