| name | deep-research |
| description | Comprehensive, citation-heavy research workflow with session state, structured outputs, and source-quality checks. Use when the user explicitly asks for deep research, a long-form report, or citation-backed analysis on a topic. |
deep-research for Codex
Read these first:
references/phase_contracts.md
references/citation_rules.md
references/quality_rubric.md
Read as needed:
references/query_generator.md
references/tool_strategy.md
references/agent_prompts.md
references/query_schema.json
examples/*.json
Use this skill when the user explicitly wants deep research, a citation-backed report, or a structured multi-phase investigation.
Codex Workflow
- Determine the mode:
- new research
- resume existing session
- status
- If the request is under-scoped, ask one compact chat-first clarification covering only the biggest unknowns.
- Create or resume a session under
RESEARCH/ using the local orchestrator helpers.
- Break the topic into subtopics and generate date-aware search queries.
- Gather sources with current official or primary references wherever possible.
- Track source quality and contradictions before synthesis.
- Write the requested output with citations and clear source provenance.
Guardrails
- Default to local execution of the research flow.
- Use live browsing for current facts and source verification.
- Prefer primary sources over summaries whenever available.
- Use sub-agents only when the user explicitly asks for team-style or parallel research, and keep each delegated subtopic bounded.
- Every key claim should have at least two supporting sources when feasible.
- If evidence is weak or contradictory, say so plainly.
Outputs
Expected workspace shape:
RESEARCH/<session_id>/
state.json
README.md
sources/
outputs/
artifacts/