| name | research |
| description | Use this skill any time the user wants to research, investigate, look up, find out about, or build understanding of a topic — including product comparisons, technical questions, best practices, "what's the state of the art on X", "what do we already know about Y", "compare A and B", or "give me a deep-dive on Z". Searches Cognee (graph + chunks + summaries) and Mem0 (user context) in parallel for prior knowledge, fills gaps via web search, returns a BLUF report with sources, and persists new findings. Distinguishes existing knowledge from new findings, flags single-source claims, and detects contradictions between stores.
|
Research Skill
Investigate a topic by searching existing knowledge stores and the web, then persist findings.
Argument Parsing
The user's message is the research topic or question. If absent, use AskUserQuestion to clarify.
Workflow
Step 1: Search Existing Knowledge
Launch two background sub-agents in parallel via the Agent tool (run_in_background: true):
Agent A — Cognee search (skip silently if Cognee MCP is not connected):
- Run three query types: GRAPH_COMPLETION (relationships), CHUNKS (facts), SUMMARIES (overviews)
top_k=15 for broad exploration
- Topic: user's question
Agent B — Mem0 search (skip silently if Mem0 MCP is not connected):
- Search Mem0 for user-related context on the topic
- Topic: user's question
Step 2: Gap Analysis and Contradiction Detection
When agents return:
- List what is already known (from Cognee + Mem0)
- Contradiction check: if Cognee and Mem0 conflict, flag explicitly and investigate which is current
- Identify gaps — questions still unanswered
- If no gaps, skip to Step 4
Step 3: Web Research
Launch a research sub-agent to fill gaps:
- Provide specific gap questions as search targets
- Agent uses WebSearch and WebFetch
- Source credibility: tag every finding with source type (official docs / engineering blog / forum post / vendor marketing). Flag findings backed by a single low-credibility source
- Agent saves new findings to Cognee (if connected) before returning
Step 4: Report
Present findings in BLUF format:
## Conclusion
[1-2 sentence answer]
## Key Findings
[Numbered list of facts/evidence]
## Sources
[Where each finding came from: Cognee / Mem0 / web URL]
## Gaps
[What remains unknown, if anything]
Add a Sources: section at the end with markdown hyperlinks if web sources are used (matches Cowork's web search citation convention).
Step 5: Persist
If Step 3 was skipped (existing knowledge sufficed), still save the synthesized conclusion to Cognee via cognify — the synthesis itself is new knowledge.
When NOT to use
- The user wants a quick factual answer that's already in Claude's training data
- The topic is clearly out of scope of any connected knowledge store and not web-searchable
- The user has already done the research and just wants implementation