Research Orchestrator

Procedure

Stage 1: SCOPE

Read research-plan.md if it exists. Ask the user:

• What is the specific research question?
• What would count as a satisfactory answer?
• How deep should the investigation go? (quick survey vs thorough review)
• Any constraints? (time range, must-include papers, specific subfield)

State the scoped question in one sentence.

Stage 2: HYPOTHESIZE

Load and follow the hypothesis-generation skill:

files_cat("/global/skills/hypothesis-generation.md")

Generate ≥5 approaches, quick-test top 3, recommend one with fallback. De-risk: test the riskiest assumption of the recommended approach first.

Stage 3: DISCOVER

Load and follow paper-discovery:

files_cat("/global/skills/paper-discovery.md")

Use the recommended approach from Stage 2 to guide search strategy.

Branch: If discovery finds <3 relevant papers:

• Broaden search terms
• Try alternative sources
• If still insufficient, consider pivoting: files_cat("/global/skills/pivot.md")

Stage 4: ANALYZE

Load and follow paper-analysis:

files_cat("/global/skills/paper-analysis.md")

Analyze top 3-5 papers from discovery. For more papers, use parallel-investigation:

files_cat("/global/skills/parallel-investigation.md")

Branch: If a paper contradicts the working hypothesis:

• This is a finding, not a problem
• Document the contradiction
• Check if hypothesis needs revision

Stage 5: SYNTHESIZE

Read all analysis notes. Identify themes, consensus, contested points, and gaps. Organize by theme, not by paper.

This is the hardest stage. Watch for regression to the mean — your synthesis should be genuinely insightful, not a summary of summaries.

Stage 6: VERIFY

Load verification-cascade for key synthesis claims:

files_cat("/global/skills/verification-cascade.md")

Apply L2+ to all claims that will appear in the final output. Apply L3 (adversarial) to the central finding.

Stage 7: CRITIQUE

Load self-critique:

files_cat("/global/skills/self-critique.md")

Evaluate the entire output. Target ≥8/10 on all dimensions. If below threshold: revise and re-critique (max 2 iterations).

Stage 8: FINALIZE (always runs, even on failure)

1. Save the research output as a workspace note
2. Update research-plan.md:
   • Move thread to Completed (or update status)
   • Add Dead Ends if any approaches were abandoned
   • Update Methodology Notes with what worked
   • Add Decision Log entry
3. Write research log entry:

   files_cat("/global/skills/research-log.md")
   

Long Run Protocol

• Checkpoint: after each stage, note the completed stage in research-plan.md
• Resume: read research-plan.md to find last completed stage, continue from next
• Exit criteria: Stage 8 completed (finalize always runs)
• If interrupted: next session's SessionStart hook will show the plan state

Output Format

The orchestrator doesn't produce a single output file. Its outputs are:

1. Discovery note (from Stage 3): research/discovery/<topic>
2. Analysis notes (from Stage 4): research/papers/<id> per paper
3. Review/synthesis note (from Stage 5-7): research/reviews/<topic>
4. Updated research-plan.md (from Stage 8)
5. Research log entry (from Stage 8)

Examples

Input: /research "How does RLHF compare to DPO for LLM alignment?"

Execution:

1. SCOPE: "Compare RLHF and DPO methods for LLM alignment — effectiveness, computational cost, failure modes"
2. HYPOTHESIZE: 5 approaches generated. Recommended: benchmark survey (search for direct comparison papers). Fallback: individual analysis of seminal papers.
3. DISCOVER: 22 papers found, 8 scored ≥4 relevance. Top 5 flagged.
4. ANALYZE: 5 papers analyzed in detail. Key finding: DPO eliminates reward model but may underperform RLHF on out-of-distribution inputs.
5. SYNTHESIZE: 3 themes — efficiency (DPO wins), robustness (RLHF wins), simplicity (DPO wins). Gap: few papers test on real-world tasks.
6. VERIFY: Central claim (DPO underperforms OOD) cross-referenced across 3 papers. Confirmed with qualification: "on specific benchmarks."
7. CRITIQUE: 8.5/10 overall. Minor issue: gap analysis could be deeper.
8. FINALIZE: review note saved, plan updated, log written.

Verification Checklist

• All 8 stages executed (or branches taken with justification)
• Stage 8 (finalize) always runs, even on failure
• Research plan updated with current state
• Research log entry written
• Self-critique score ≥8/10

When Stuck

• At any stage: consider pivoting. Load pivot skill.
• If the question is too broad: narrow the scope at Stage 1
• If no papers exist: this might be an open problem — document that as a finding
• If synthesis is weak: you probably didn't analyze enough papers — go back to Stage 4

Integration

• This is the top-level skill. It composes all others.
• Entry point: user invokes /research <question>
• Resume: reads research-plan.md for state, continues from last checkpoint