create-literature-report

name: create-literature-report version: 2.0.0 description: > Reproducible scientific literature research report with citation-verified, hallucination-free synthesis. 6-phase pipeline composing Markov novelty writing, parallel web research, and zero-tolerance citation verification. last_updated: 2026-03-19 activation: See ACTIVATION-CHECKLIST.md invocation:

• "create literature report on [topic]"
• "research report on [topic]"
• "literature survey for [topic]" revision_notes: > v2.0.0: Decomposed monolith into protocols + prompts + state schema. Added ACTIVATION-CHECKLIST, agent prompt templates, WebFetch failure handling, deduplication algorithm, FORCE_STOP protocol, workspace structure. Addresses 19 architectural gaps from 3-reviewer audit. metadata: category: research tags: [literature, citations, web-search, synthesis, hallucination-free] relations: compose_with:
  • prd-update
  • kg-sync depend_on:
  • fetch-docs similar_to: [] belong_to: []

create-literature-report v2.0.0

Zero-hallucination scientific literature synthesis in 6 phases.

Quick Start

1. Run ACTIVATION-CHECKLIST.md pre-flight
2. Execute phases 0→5 sequentially
3. State file enables crash recovery at any sub-step

Parameters

Architecture: 6-Phase Pipeline

Phase 0: CAPTURE & ALIGN         (~5 min)
    ↓
Phase 1: GATHER (parallel)       (~15 min)
    ↓
Phase 2: SYNTHESIZE              (~15 min)
    ↓
Phase 3: VERIFY (parallel)       (~10 min)
    ↓
Phase 4: REVIEW (iterative)      (~10 min, optional)
    ↓
Phase 5: DELIVER                 (~5 min)

Total time budget: <60 minutes for 50-paper report.

Phase Protocols

Each phase has a dedicated protocol file with exact steps:

Agent Prompt Templates

State Management

• Schema: state/literature-report-state.schema.json
• Example: state/example-state.json
• Location: state/literature-report-{topic_slug}-state.json
• Sub-step tracking enables precise crash recovery

Workspace Structure

state/
  literature-report-{topic_slug}-state.json
workspace/literature-report-{topic_slug}/
  phase-1-gather/
    local-inventory.json       # Structured inventory from local agent
    web-research-a.json        # Papers from research agent A
    web-research-b.json        # Papers from research agent B
    deduplicated-papers.json   # Final merged, deduplicated list
  phase-3-verify/
    verification-batch-1.json  # Results from verifier agent 1
    verification-batch-2.json  # Results from verifier agent 2
    verification-batch-3.json  # Results from verifier agent 3
    corrections-log.json       # All corrections applied
  phase-4-review/
    iter-1/
      review-domain.md         # Domain expert feedback
      review-novelty.md        # Novelty assessor feedback
      synthesis.md             # L2 synthesis
      verdict.md               # L1 verdict

Convergence Criteria

FORCE_STOP Protocol

If any phase exceeds its time budget or hits 3 consecutive failures:

1. Save current state with "force_stopped": true, "reason": "..."
2. Git commit all work so far
3. Report to user: phase, substep, failure reason, recommended action
4. Do NOT continue to next phase
5. On resume: read state, fix the blocking issue, restart current phase

Anti-Patterns (BANNED)

Eval Criteria

Quantitative (automated)

Qualitative (reviewer agents)

Regression test

Run on biomedical-agentic-ai topic with original 27 seeds. Expected: 0 hallucinations, ≥55 verified papers, ≥3 novel insights.

Sub-Skill Dependencies

Note: These skills provide principles that are inlined into this skill's protocol files. This skill is self-contained — it does NOT require the sci-llm-writer repo to be present at runtime.

Version History