Literature review

<SUBAGENT-STOP> If you were dispatched as a subagent to execute a specific review task, skip this skill. Do the task and return structured results immediately. </SUBAGENT-STOP> <HARD-GATE> Do NOT launch searches or dispatch subagents until you have written the plan. Write the plan and continue immediately. Do not ask for confirmation unless the user explicitly requested plan review. </HARD-GATE>

Literature Review Skill — Parallel Search + Evidence Synthesis

Workflow Overview

PLAN → GATHER (parallel researcher subagents) → SYNTHESIZE → CITE (verifier) → VERIFY (reviewer) → DELIVER

---

Step 1: Plan

Derive a short slug from the topic (lowercase, hyphens, no filler words, ≤5 words). Use this slug for all files.

Create .research-loop/sessions/<slug>/plan.md with:

• Key questions the review aims to answer
• Search angles (3–6 distinct angles for parallel dispatch)
• Source types to search: papers, web docs, repos, benchmarks, conference proceedings
• Time period / recency window
• Expected sections for the final review
• Task ledger tracking each angle
• Verification log

Write the plan and continue immediately. Summarize briefly. Do not ask for confirmation.

---

Step 2: Gather Evidence

Parallel search across multiple angles

Dispatch researcher subagents (.claude/agents/researcher.md) for each search angle. For a typical literature review, use 3–4 subagents in parallel:

Agent	Task
researcher	Search academic papers: foundational works, recent advances, surveys
researcher	Search official documentation, benchmarks, and implementation repos
researcher	Search for field disagreements, unresolved debates, open problems
researcher (if broad scope)	Search adjacent fields, cross-domain applications, alternative approaches

Each researcher writes to .research-loop/sessions/<slug>/findings-<angle>.md.

Evidence table with quality scoring

Each researcher must produce an evidence table with:

#	Source	URL	Key claim	Type	Quality
1	...	...	...	primary / secondary / self-reported	A / B / C

Source quality tiers

Tier	Description	Examples
A (Highest)	Peer-reviewed papers, official documentation, verified benchmarks, primary datasets, government filings	arXiv proceedings, IEEE/ACM, official docs
B (Good)	Reputable secondary sources, expert technical blogs, well-cited surveys, established trade publications	Distill.pub, high-quality blog posts
C (Accept with caveats)	Undated posts, content aggregators, social media with primary links, vendor claims without independent verification	Listicles, Medium posts without citations
Reject	No author + no date, AI-generated content without primary backing, anonymous claims	—

When initial results skew toward low-quality sources, re-search with domain filters targeting authoritative domains (.edu, .gov, arXiv, ACL, NeurIPS, etc.).

Context hygiene for researchers

• Write findings progressively to files. Do not accumulate full page contents in working memory.
• Use includeContent: true for top candidates only; triage by title/snippet first.
• Triage 10+ results by title before fetching full content.
• Track assigned questions explicitly: mark each as done, blocked, or needs follow-up.

---

Step 3: Synthesize

Write the synthesis yourself. Do not delegate to subagents.

Save to .research-loop/sessions/<slug>/draft.md.

Structure:

• Executive Summary — 2–3 paragraph overview of findings
• Topic Clusters — Group findings by thematic clusters, not by source
• Consensus Points — What the literature agrees on
• Disagreements — Where experts diverge, with steel-manned arguments for each side
• Research Gaps — Explicit gaps identified in the literature; unanswered questions
• Open Questions — What remains unresolved and why it matters

Synthesis rules:

• Separate consensus, disagreements, and open questions clearly
• When useful, propose concrete next experiments or follow-up reading
• Before finishing the draft, sweep every strong claim against the verification log
• Downgrade anything that is inferred or single-source critical

---

Step 4: Cite

Dispatch the verifier agent (.claude/agents/verifier.md) to add inline citations and verify every source URL.

The verifier will:

• Anchor every factual claim to a numbered source citation
• Verify every URL resolves and contains the claimed content
• Remove unsourced factual claims
• Build a unified Sources section
• Deduplicate sources across multiple research files
• Enforce: every [N] in body maps to Sources, every Sources entry is cited at least once

Save the cited draft to .research-loop/sessions/<slug>/draft-cited.md.

---

Step 5: Verify

Dispatch the reviewer agent (.claude/agents/reviewer.md) to check the cited draft for:

• Unsupported claims
• Logical gaps
• Zombie sections (content surviving from earlier drafts without support)
• Single-source critical findings
• Overstated confidence

The reviewer classifies issues as:

• FATAL — must fix before delivery
• MAJOR — note in Open Questions
• MINOR — accept as-is

Fix FATAL issues before delivery. If FATAL issues were found, run one more verification pass after the fixes.

---

Step 6: Deliver

Save the final literature review to .research-loop/sessions/<slug>/report.md.

Write a provenance sidecar to .research-loop/sessions/<slug>/provenance.md:

# Provenance: Literature Review — [topic]

- **Date:** [date]
- **Search angles:** [list of angles]
- **Sources consulted:** [count and/or list]
- **Sources accepted:** [count]
- **Sources rejected:** [count] — [reasons]
- **Verification:** [PASS / PASS WITH NOTES / BLOCKED]
- **Plan:** .research-loop/sessions/<slug>/plan.md
- **Research files:** [files used]

Before responding, verify on disk that both report.md and provenance.md exist.

---

Subagent Reference

Agent	When dispatched
researcher	Step 2 — evidence gathering across parallel angles
verifier	Step 4 — citation verification and URL checking
reviewer	Step 5 — review pass on cited draft

All subagent definitions live in .claude/agents/.