Do you want to conduct deep web research?

Step 1: Decompose the Question

Before searching anything, break the user's research request into discrete sub-questions. Output them as a numbered list.

Rules:

• Each sub-question should be independently answerable.
• If the user's request is already a single, specific question, you still need at least 2-3 angles to investigate (e.g., "What is X?" becomes: "What is X?", "How does X compare to alternatives?", "What are the limitations or criticisms of X?").
• Symmetrical Baseline sub-question (required). Per the de-anchoring rule in analytical_integrity skill, you MUST add at least one sub-question that abstracts the core phenomenon from the specific actors and checks if opposing parties or the industry at large are doing the same thing. All rules in that skill file also apply throughout the research workflow.
• Always add at least one "Contextual Sweep" sub-question. This question must look for concurrent events, lawsuits, earnings calls, or controversies involving the primary entities around the relevant timeframe that might provide hidden motives or broader context (e.g., "What other major news involving [Entity] happened in [Month/Year]?").
• If you're unsure whether the user wants a shallow answer or deep investigation, default to deep investigation.

I. Step 2: Search Loop

For each sub-question, gather information from at least 3 independent sources.

Rules:

• "Independent" means different websites, authors, or publications. Three paragraphs from the same blog post do not count as three sources.
• One source is only sufficient if it is an authoritative primary source (e.g., official documentation, the actual API reference, a court filing, a peer-reviewed paper) that definitively and completely answers the sub-question with no ambiguity.
• If the source is a blog, video, opinion piece, forum post, or news article, you must find at least 2 more sources to corroborate.
• For the Contextual Sweep sub-question, use broad time-bound queries (e.g., "[Entity] news [Month Year]", "[Entity] lawsuit [Date]") to catch overlapping events that weren't explicitly mentioned in the user's prompt.
• Use a mix of tools: web search, URL reads, video transcripts, visible_browser MCP tools, etc. Don't rely on a single tool type.

What to record for each source:

• The URL or source identifier
• A brief summary of what it said relevant to the sub-question
• Whether you consider it a primary/authoritative source or a secondary/opinion source

II. Step 3: Handle Source Failures

If a source is blocked, a transcript fails, a page returns garbage, or an API errors out:

• Do NOT skip the sub-question. "I couldn't access it" is never a valid reason to move on.
• Try a different tool (e.g., if read_url_content fails, use visible_browser MCP tools).
• Try a different search query.
• Try a different source entirely.
• Only flag a sub-question as "unable to resolve" after at least 3 genuinely different attempts to find information on it.

III. Step 4: Cross-Reference

After gathering sources for all sub-questions, compare findings across sources.

Rules:

• If sources agree, note the consensus.
• If sources contradict each other, flag the contradiction explicitly. Do NOT silently resolve it by picking whichever source you read first.
• When contradictions exist, present both sides with their sources and let the user decide, or explain which source is more authoritative and why.

IV. Step 5: Gaps Check

Review your sub-question list from Step 1. For each one, verify:

• Does it have at least one substantiated answer with a cited source?
• If not, attempt alternative approaches: different search terms, Deep Research, asking the user for clarification.
• Any sub-questions that remain unanswered after exhausting alternatives must be explicitly flagged as unresolved in your final report.

V. Step 6: Report

Compile your findings into a structured artifact (markdown file in the project directory).

The report must include:

• The original question/request
• The sub-questions you decomposed it into
• Findings for each sub-question, with inline source citations
• Any contradictions found between sources
• Any sub-questions that remain unresolved
• A summary/conclusion synthesizing the key findings

You may NOT declare research "done" until:

• Every sub-question from Step 1 has at least one substantiated answer OR is explicitly flagged as unresolved.
• All contradictions are surfaced, not hidden.
• The report artifact is written and saved.

Only after the report is complete, give the user a brief verbal summary highlighting the key findings and any open questions.

• 01_research_findings.md — The comprehensive research report.