Research Automation Skill
Automate your research process from literature survey to paper draft with a structured pipeline.
Fully automated research pipeline from literature survey to paper draft.
Overview
The /research command orchestrates an 8-phase pipeline driven by a Master Agent (you) that acts as a "Research PI" — making decisions, reviewing outputs, debugging failures, and iterating until completion.
Architecture
Master Agent (you, Opus) — full-context decision maker
├── Phase 1: Literature Survey ← Worker: literature-searcher
├── Phase 2: SOTA Codebase Discovery ← Worker: sota-finder
├── Phase 3: Idea Generation ← Master does this (needs global context)
├── Phase 4: Experiment Design ← Master does this
├── Phase 5: Baseline Reproduction ← Workers: env-setup + experiment-runner
├── Phase 6: Experiments ← Worker: experiment-runner + Master reviews
├── Phase 7: Analysis ← Master + Worker: data-analyst
└── Phase 8: Paper Draft ← Master does this
Key Principles
- Failure is information, not disaster — every phase outputs either success results or valuable failure analysis
- Iterate like a human researcher — no hard retry limits; Master Agent judges when to debug, pivot, or stop
- Three-layer correctness — Information correctness (Phase 1-2) → Code correctness (Phase 5-6) → Conclusion correctness (Phase 7-8)
- Never fabricate — results.json is write-protected by PreToolUse hook; report failures honestly
- Cross-session resumable — state.json + reasoning.md enable any new session to continue
State Management
All state lives in .research/state.json. See references/phase-definitions.md for the full schema.
Key operations:
from scripts.lib.research_utils import StateManager, create_research_dirs
sm = StateManager(".research")
state = sm.init("your topic", depth="full")
state = sm.advance_phase(state, "phase1_literature", "phase2_sota")
context = sm.get_resume_context(state) # For --resume
Phase Transition Gates
Each phase must pass a quality gate before advancing. Gates are defined in references/phase-definitions.md.
If a gate fails: attempt auto-fix (max 2 tries) → still fails → mark phase as gate_failed → report in final output what was achieved.
Three Defense Layers
- Micro-experiment verification (after code changes, before full training) — see
references/micro-experiment.md - Reproduction comparison (baseline vs paper values, <5% = excellent, 5-15% = acceptable, >15% = investigate)
- Anti-fabrication (PreToolUse hook blocks Write/Edit to results.json — mechanical enforcement)
Error Handling
Errors are classified into 6 categories with typed recovery routes. See references/error-taxonomy.md.
API Reference
Verified API endpoints for paper search and SOTA discovery. See references/api-reference.md.
Related Components
- Command:
/research— main entry point - Agents:
sota-finder,env-setup,experiment-runner,opportunity-scorer - Reused agents:
literature-reviewer,data-analyst,architect,code-reviewer - Reused skills:
ml-paper-writing,paper-self-review,writing-anti-ai,citation-verification