메뉴
IterDRAG iterative retrieval-augmented generation pattern — search, summarize, reflect loop for deep research tasks
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
| name | deep-research |
| description | IterDRAG iterative retrieval-augmented generation pattern — search, summarize, reflect loop for deep research tasks |
| tags | ["research","search","rag","deep-research","analysis"] |
When assigned a research task, follow this iterative loop until the answer is comprehensive and well-sourced.
QUERY → DECOMPOSE → SEARCH → SUMMARIZE → REFLECT → (gaps?) → SEARCH again → REPORT
Stop when: all sub-questions have authoritative answers, OR 3 iterations complete, OR new searches return no new information.
Before searching, break the research question into 3–5 independently searchable sub-questions that together cover the full topic.
Example — "What are best practices for Connect-RPC in Go?":
For each sub-question:
After searching all sub-questions:
Ask:
If gaps found → formulate new targeted queries and return to Phase 2. If confident → proceed to Phase 5.
# [Topic] — Research Report
## Summary
[2–3 sentences: key findings]
## Findings
### [Sub-question 1]
[Answer with inline citations]
### [Sub-question 2]
[Answer with inline citations]
## Key Recommendations
- [Actionable recommendation 1]
- [Actionable recommendation 2]
## Sources
- [URL or doc name] — [why relevant]
## Gaps and Caveats
- [What remains uncertain or needs verification]
| Mode | When to use |
|---|---|
| Local (Ollama) | Private data, offline, cost-sensitive. Models: llama3, mistral, deepseek-r1 |
| Cloud | High accuracy on complex topics. Models: claude-sonnet-4-6, gpt-4o |
Search providers (no key required): DuckDuckGo instant answers. Search providers (API key): Brave Search API, Tavily API, SerpAPI.
Good output:
Poor output:
Use whenever the user wants to run, plan, kick off, iterate on, or close out a Victoria training version (e.g. "let's start V200", "next training run", "kick off the next iteration", "what should we try next for Victoria", "ship V###", "log this training run"). Codifies the continual-improvement loop — read the latest entry in omega/nodes/victoria/training_log/, propose the next version's hypothesis, run the gates, write the new V###.md entry, update the high-water table, commit, and push.
Structured research pipeline for Omega — takes an idea or hypothesis, searches for evidence, compares to existing Omega capabilities, and delivers a verdict with implementation recommendations. Use when asked to /research a topic, evaluate a new signal source, assess an external tool/library, or investigate trading strategies.
Go coding standards, error handling, concurrency, and Connect-RPC conventions for the Omega project
Buf conventions, proto3 best practices, schema evolution rules, and Connect-ES v2 frontend patterns for Omega
Go table-driven tests, Python pytest patterns, React Testing Library, property-based testing, and integration test conventions for Omega