Research

Structured web research with persistent, queryable storage. All research persists in ~/Claude-Workspaces/.

Workflow

1. •

   Scope — Clarify what the user needs. Classify:

   • •quick-lookup: Single question, 1-2 searches, inline answer
   • •deep-research: Multi-faceted, 2-4 parallel subagents
2. •

   Create session — Run:

   bash

   python3 "$SKILL_DIR/scripts/research_db.py" create-session --type <type> --topic "Topic Name"
   

   Capture the returned session_dir from JSON output.

3. •

   Check existing research — Query the DB first:

   bash

   python3 "$SKILL_DIR/scripts/research_db.py" search "relevant term"
   

   Reuse and build on prior findings.

4. •

   Execute search:

   Quick lookup: Search directly with WebSearch/WebFetch. Answer inline.

   Deep research: Decompose into 2-4 independent facets. Dispatch ALL as subagents via Task tool in ONE message for true parallelism:

   code

   Task(subagent_type="research", prompt="Research facet: <specific angle>. Return JSON: {findings, sources: [{url, title, accessed}], confidence: high|medium|low}")
   Task(subagent_type="research", prompt="Research facet: <another angle>. Return JSON: ...")
   

   Each subagent:

   • •Gets one specific facet/angle
   • •Returns structured JSON (findings, sources, confidence)
   • •Does NOT write files — parent synthesizes
5. •

   Synthesize — Merge results. Cross-reference claims across sources. Flag conflicts. Assign confidence:

   • •high: 3+ agreeing sources
   • •medium: 2 sources or minor conflicts
   • •low: single source or significant conflicts
6. •

   Persist — Save to DB and session directory:

   bash

   python3 "$SKILL_DIR/scripts/research_db.py" add \
     --topic "Topic" --query "Original question" --summary "Findings" \
     --raw-findings "Detailed notes" \
     --sources '[{"url":"...","title":"...","accessed":"2026-02-09"}]' \
     --tags "tag1,tag2" --confidence high \
     --session-dir "$SESSION_DIR" --session-type deep-research
   

7. •

   Deliver — Concise summary with confidence markers and source links. No lengthy reports unless asked.

Path Resolution

Locate this SKILL.md on disk, then use its parent directory as SKILL_DIR. The script is at $SKILL_DIR/scripts/research_db.py.

DB location: ~/Claude-Workspaces/research.db (override with RESEARCH_WORKSPACE env var).

Failure Handling

• •If WebSearch/WebFetch fails, log the failure, try alternate queries
• •Report partial results with low confidence rather than failing silently
• •Save research even if findings are negative ("X does not exist" is valuable)

Research Quality Rules

• •Never present single-source findings as fact — flag as low confidence
• •Prefer primary sources over aggregator/blog posts
• •Include access dates in source metadata
• •When sources conflict, present both sides
• •For technical topics, verify against official documentation