deep-research

# Deep Research > If you do not know, research first. Do not pretend to know. --- ## WHEN TRIGGERED - EXECUTE IMMEDIATELY If topic is provided, start with Step 1. If not, ask what to research. --- ## Workflow ### Step 1: Scope the research **Type**: prompt From `$ARGUMENTS`, define: - **Topic**: what exactly to research - **Key questions**: 3-5 specific questions to answer - **Boundaries**: what is in scope vs out of scope - **Depth**: lite or deep mode If topic is vague, ask ONE clarifying question before proceeding. ### Step 2: Plan retrieval (Deep mode only) **Type**: prompt For deep mode, design a retrieval plan: - Assign each question to a search strategy (web, academic, documentation, code) - Identify 2-3 source types per question (official docs, blog posts, papers, repos) - Plan parallel agent assignments ### Step 3: Execute research **Type**: script + prompt **Lite mode:** - Use WebSearch for each key question - WebFetch top 2-3 results per question - Extract relevant evidence and note sources **Deep mode:** - Spawn 2-3 research agents in parallel using Agent tool - Each agent handles a subset of questions - Each agent returns: findings, sources, confidence level, gaps ```bash # Ensure output directory exists mkdir -p "$OUTPUT_DIR" ``` ### Step 4: Triangulate sources **Type**: review For each finding: - Check if 2+ independent sources confirm it - Mark single-source claims as "unverified" - Flag contradictions between sources - Note recency — prefer sources from last 12 months ### Step 5: Synthesize and output **Type**: generate Write findings to `{output_dir}/`: **Main report** (`{topic-slug}-research.md`): ```markdown # Research: {topic} Date: {date} Depth: {lite/deep} ## Key Findings 1. {finding with source} 2. {finding with source} ## Detailed Analysis ### {Question 1} {evidence and analysis} ### {Question 2} {evidence and analysis} ## Sources - [{title}]({url}) — {relevance note} ## Gaps and Uncertainties - {what we couldn't confirm} ``` Report to user: ``` Research complete: {topic} Depth: {lite/deep} Questions answered: {N}/{total} Sources consulted: {N} Output: {output_dir}/{filename} ``` --- ## Edge Cases **No relevant results**: Report "Limited sources found for this topic" and suggest alternative search terms or narrower scope. **Conflicting sources**: Present both sides clearly. Do not pick a winner without evidence. Mark as "contested." **Rate-limited search**: If WebSearch fails, wait and retry once. If still failing, use cached/known information and mark as "incomplete." **Very broad topic**: Suggest narrowing. "This topic is too broad for {depth} mode. Suggested sub-topics: ..." **Sensitive/controversial topics**: Present facts neutrally. Cite all sides. Do not take positions. --- ## Integration Points - **content-draft**: Research → content pipeline. Research output feeds into blog posts, presentations. - **data-analyst**: If research involves data files, delegate to data-analyst for statistical analysis. - **meeting-notes**: Research findings can be prepared as meeting discussion material. - **memory-capture**: Promote key research findings to durable memory if they inform ongoing project decisions. - **morning-briefing**: Scheduled research updates can feed into daily briefings via Cloud Tasks.