claude-research

claude-research

Drives the Research feature on claude.ai using the user's already-authenticated Chrome session via Chrome DevTools MCP (Playwright MCP as fallback), then returns the produced report.

When to use

Invoke when the user wants to:

• Run a Research / deep-research query on claude.ai and bring the result back into the current Claude Code session.
• Kick off a long-running claude.ai investigation while continuing other work locally.
• Re-open a previous claude.ai Research conversation to extract its output.
• Just verify the setup (MCP attach + login) without submitting a query — treat that as the preflight-only path (steps 1–5, then report).

Do NOT use for ordinary (non-Research) chats, scraping non-claude.ai sites, or one-off questions the local model can answer directly.

Tool fallback order

1. Primary: Chrome DevTools MCP — tools named mcp__chrome-devtools__*.
2. Fallback: Playwright MCP — tools named mcp__playwright__*, configured to connect over CDP.

Pick one MCP at the start and stay with it. Don't mix tool families mid-flow.

Clarification handling — default is auto-answer

claude.ai's Research mode almost always asks 1–3 follow-up questions before kicking off the actual deep-research run (scope, time period, depth, regions, etc.). The whole point of this skill is to fire-and-forget a research run from Claude Code, so the default is to auto-answer those clarifications and let the run start unattended.

Auto-answer rules:

• Read what claude.ai is asking.
• Reply with the most natural answer the original user query implies. Use the original query's words and intent — don't introduce scope, constraints, or preferences the user didn't ask for.
• When the original query is silent on a point, defer to claude.ai's judgment. Short answers like "No preference — use your best judgment", "Whatever you think is most useful", or "Up to you, just proceed" are fine.
• If claude.ai shows a "Start research" / "Skip clarifications" / "Use defaults" button, prefer clicking that over composing a reply.
• Loop: if claude.ai asks again after your reply, answer again. Hard cap at 3 clarification rounds — if it's still asking after 3 replies, something is off; bail to the user with what you saw.

Bail out of auto-answer and surface the question to the user when:

• The user invoked the skill with explicit "ask me", "interactive", "wait for my input", or "check with me on the plan" framing.
• The clarification asks for a fact that's genuinely not derivable from the original query AND would materially change the research direction (e.g., "Which of these three companies should I focus on?" when the user didn't name any). Picking arbitrarily here would silently change what the user gets back.

When in doubt, lean toward auto-answering with "use your best judgment" — claude.ai's defaults are usually fine, and the user can always re-run with a more specific query.

Preflight (always run these first)

1. Detect a browser MCP

Look at the registered tools in this session:

• mcp__chrome-devtools__* present → use Chrome DevTools MCP.
• Otherwise mcp__playwright__* present → use Playwright MCP.
• Neither → stop. Tell the user to install one (point at README.md).

State which MCP you picked before doing anything else.

2. Confirm the MCP is in attach mode, not launch mode

chrome-devtools-mcp has two modes:

• Attach mode — registered with --autoConnect (Chrome 144+) or --browser-url=http://127.0.0.1:9222. Drives the user's real, logged-in Chrome.
• Launch mode (default) — spins up its own Chrome with a throwaway profile at ~/.cache/chrome-devtools-mcp/chrome-profile-stable. No logins. Wrong for this skill.

Sanity-check which mode is active before navigating to claude.ai. The simplest signal: navigate to chrome://version and read Profile Path from the accessibility snapshot. If it looks like the throwaway cache dir, stop and tell the user:

• The MCP is in launch mode.
• Point them at the README's "Verified setup" section (--autoConnect) or "Option B — --browser-url".
• Suggest claude mcp list to confirm current flags.

Do NOT log in inside the throwaway profile — that defeats the point.

3. Navigate to claude.ai

Open https://claude.ai in the attached browser.

4. Verify login

Read the page via the MCP's accessibility-tree / snapshot tool:

• Logged in signals: sidebar with chat history, the composer (placeholder Type / for skills or similar), a "+" button next to the composer, the user's avatar.
• Not logged in signals: a "Log in" / "Sign up" button, marketing homepage layout, an email/password form.

If signals are ambiguous (interstitial loading), wait a few seconds and re-snapshot once before deciding. If not logged in, stop and ask the user to log in in their Chrome window — do not attempt to log in yourself.

5. (Preflight-only path)

If the user only asked for a preflight, report and stop:

• Which MCP, which mode, which profile path you observed.
• Current URL.
• Logged in / not logged in verdict + the signal you matched.

Otherwise continue to step 6.

Running a Research query

6. Open the composer's + menu

In the claude.ai composer there is a small + button below or beside the message input. Click it. A menu opens with entries that include:

Add files or photos
Take a screenshot
Add to project        >
Add from GitHub
---
Skills                >
Add connectors
---
Research
Web search            ✓   (checkmark when active)
Use style             >

These names are observed UI labels, not selectors. Find them in the accessibility tree by role + accessible name, not by CSS classes or DOM positions — claude.ai's markup changes.

If the + button is missing, the composer might be in a different state (a chat already open, mobile layout, etc.). Try clicking "New chat" first to get a fresh composer, then re-check.

7. Enable Research mode

Click the Research entry in the menu. After clicking:

• Re-snapshot the menu (or the composer if the menu closed) and confirm Research is now toggled on. The checkmark / highlighted state appears on the active item, mirroring how "Web search ✓" indicates an active toggle.
• If toggling Research disabled "Web search" (they may be mutually exclusive in claude.ai's UI), that's expected — note it in the final report but don't try to re-enable Web search.
• If Research doesn't appear in the menu at all, the user's account may not have Research access. Stop and report what entries you did see.

Close the menu (Escape or click outside) before composing.

8. Compose and submit

• Click into the message input.
• Type the user's query verbatim. Do not paraphrase, expand, or "improve" it.
• Submit (Enter, or the send button — whichever is exposed by the accessibility tree).

claude.ai's Research mode will almost always show one or more clarification prompts before kicking off the actual run. Handle them per the Clarification handling section above:

• Default: auto-answer (deriving the reply from the original user query, deferring to claude.ai's defaults when the original is silent).
• Override (caller said "ask me", "interactive", etc.) or genuinely-ambiguous direction-changing question: stop and surface to the user.
• Loop until claude.ai stops asking clarifications and the actual research starts streaming. Cap at 3 rounds.

Note in the final report (step 11) how many clarification rounds happened and a one-line summary of what you answered, so the user can audit your auto-replies.

9. Wait for completion

Research queries take 5–30+ minutes. While waiting:

• Poll every 30–60 seconds via accessibility snapshots. Suitable completion signals: streaming indicator disappears, a "Research complete" / "Finished" marker appears, the report renders fully, "Stop" button becomes "Regenerate".
• Do not navigate away, close the tab, or trigger other actions in that tab.
• If the user asked the skill to run in the background: return immediately with the conversation URL and tell the caller to re-invoke the skill with that URL once they want to pick up the result.

10. Extract the result

Once complete:

• Capture the rendered report from the DOM / accessibility tree as text (Markdown if claude.ai exposes it; otherwise plain text). Never substitute a screenshot — the user wants the report content, not a picture.
• Expand any collapsible sections (citations, plan steps, source lists) before reading, so the full body is captured.
• Capture the conversation's claude.ai URL from the address bar.

11. Return

To the caller:

• Full report text.
• claude.ai conversation URL.
• One-line summary of elapsed time and any non-obvious events (e.g. "Research mode was enabled; Web search was deactivated automatically. Took 12 minutes.").
• Clarification audit — N rounds, plus the question(s) claude.ai asked and the reply you gave. Keep this brief (one line per round) so the user can spot if you auto-answered something they'd have answered differently.

Resuming an existing Research conversation

If the user provides a claude.ai conversation URL instead of a fresh query:

• Navigate directly to that URL (skip steps 6–8).
• If the query is still running, jump to step 9 (wait).
• If already complete, jump to step 10 (extract).

Failure modes — surface, don't paper over

• MCP in launch mode → stop at step 2; point at README's Verified setup section.
• Not logged in → stop at step 4; ask the user to log in.
• + button or Research entry not found → stop at step 6/7; report what menu entries you did see.
• Account lacks Research access → stop; surface claude.ai's exact message.
• Chrome not reachable on 9222 (Option B users only) → stop; point at README's launch command.
• Query rejected (rate limit, content policy) → surface claude.ai's exact message. Don't retry silently.

What this skill is NOT for

• Ordinary (non-Research) chats — use the API or a different skill.
• Bypassing claude.ai auth, rate limits, or content policies.
• Reading other users' conversations.
• Disabling or re-enabling other composer toggles (Web search, Style, etc.) unless the user explicitly asks.