Literature Review

Literature Review

Conduct deep literature reviews through multi-perspective dialogue and systematic search.

Input

• $0 — Research topic or question
• $1 — Optional: specific focus or angle

References

• Multi-perspective dialogue prompts (STORM): ~/.claude/skills/literature-review/references/dialogue-prompts.md
• Literature review workflow (AgentLaboratory): ~/.claude/skills/literature-review/references/review-workflow.md

Scripts (from literature-search skill)

# Search Semantic Scholar
python ~/.claude/skills/deep-research/scripts/search_semantic_scholar.py --query "topic" --max-results 20

# Search OpenAlex
python ~/.claude/skills/literature-search/scripts/search_openalex.py --query "topic" --max-results 20

# Search arXiv
python ~/.claude/skills/deep-research/scripts/search_arxiv.py --query "topic" --max-results 10

Workflow

Step 1: Generate Expert Personas (from STORM)

Given the topic, create 3-5 diverse expert personas:

• Each represents a different perspective, role, or research angle
• Example: "ML systems researcher focused on efficiency", "Theoretical statistician concerned with guarantees"
• Use the persona generation prompts from references

Step 2: Multi-Perspective Dialogue

For each persona, simulate a multi-turn Q&A conversation: 1. Persona asks a question from their unique angle 2. Generate search queries from the question 3. Search literature using the search scripts 4. Synthesize an answer grounded in retrieved papers with inline citations 5. Record the dialogue turn with search results 6. Repeat for 3-5 turns per persona 7. End when persona says "Thank you so much for your help!"

Step 3: Synthesize Knowledge

• Combine all persona conversations into a unified knowledge base
• Remove redundancy across personas
• Organize by theme/subtopic
• Generate an outline based on the collected information

Step 4: Generate Literature Review

• Write a structured review organized by the generated outline
• Every claim must be supported by a citation
• Include a summary table of key papers (method, contribution, limitations)

Output

A structured literature review with: 1. Outline — Hierarchical topic structure 2. Per-section summaries — Each grounded in retrieved papers 3. Paper database — Structured entries for all reviewed papers 4. Knowledge gaps — Identified areas needing further investigation

Rules

• Every sentence in the review must be supported by gathered information
• If information is not found, explicitly state the gap
• Cite broadly — cover diverse approaches, not just the most popular
• Include recent papers (last 2-3 years) alongside foundational work
• Use inline citations: "Smith et al. [1] propose..."

Related Skills

• Upstream: literature-search, deep-research
• Downstream: related-work-writing, research-planning
• See also: survey-generation

Literature Review Dialogue Prompts Reference

Extracted from STORM (Stanford).

1. Persona Generation

Find Related Topics

I'm writing a Wikipedia page for a topic mentioned below. Please identify and recommend some Wikipedia pages on closely related subjects. I'm looking for examples that provide insights into interesting aspects commonly associated with this topic, or examples that help me understand the typical content and structure included in Wikipedia pages for similar topics.
Please list the urls in separate lines.

Generate Expert Personas

You need to select a group of Wikipedia editors who will work together to create a comprehensive article on the topic. Each of them represents a different perspective, role, or affiliation related to this topic. You can use other Wikipedia pages of related topics for inspiration. For each editor, add a description of what they will focus on.
Give your answer in the following format: 1. short summary of editor 1: description\n2. short summary of editor 2: description\n...

Generate Focused Experts

You need to select a group of speakers who will be suitable to have roundtable discussion on the [topic] of specific [focus].
You may consider inviting speakers having opposite stands on the topic; speakers representing different interest parties.
Ensure that the selected speakers are directly connected to the specific context and scenario provided.
For each speaker, add a description of their interests and what they will focus on during the discussion.
Strictly follow format below:
1. [speaker 1 role]: [speaker 1 short description]
2. [speaker 2 role]: [speaker 2 short description]

2. Multi-Turn Q&A Dialogue

Ask Question (basic)

You are an experienced Wikipedia writer. You are chatting with an expert to get information for the topic you want to contribute. Ask good questions to get more useful information relevant to the topic.
When you have no more question to ask, say "Thank you so much for your help!" to end the conversation.
Please only ask a question at a time and don't ask what you have asked before. Your questions should be related to the topic you want to write.

Ask Question with Persona

You are an experienced Wikipedia writer and want to edit a specific page. Besides your identity as a Wikipedia writer, you have specific focus when researching the topic.
Now, you are chatting with an expert to get information. Ask good questions to get more useful information.
When you have no more question to ask, say "Thank you so much for your help!" to end the conversation.
Please only ask a question at a time and don't ask what you have asked before. Your questions should be related to the topic you want to write.

Convert Question to Search Queries

You want to answer the question using Google search. What do you type in the search box?
Write the queries you will use in the following format:
- query 1
- query 2
...
- query n

Answer Question (grounded in search results)

You are an expert who can use information effectively. You are chatting with a Wikipedia writer who wants to write a Wikipedia page on topic you know. You have gathered the related information and will now use the information to form a response.
Make your response as informative as possible, ensuring that every sentence is supported by the gathered information. If the [gathered information] is not directly related to the [topic] or [question], provide the most relevant answer based on the available information. If no appropriate answer can be formulated, respond with, "I cannot answer this question based on the available information," and explain any limitations or gaps.

3. Grounded Discussion Focus

Your job is to find next discussion focus in a roundtable conversation. You will be given previous conversation summary and some information that might assist you discover new discussion focus.
Note that the new discussion focus should bring new angle and perspective to the discussion and avoid repetition. The new discussion focus should be grounded on the available information and push the boundaries of the current discussion for broader exploration.
The new discussion focus should have natural flow from last utterance in the conversation.
Use [1][2] in line to ground your question.

4. Outline Generation

Initial Outline

Write an outline for a Wikipedia page.
Here is the format of your writing:
1. Use "#" Title" to indicate section title, "##" Title" to indicate subsection title, "###" Title" to indicate subsubsection title, and so on.
2. Do not include other information.
3. Do not include topic name itself in the outline.

Outline Refinement from Conversations

Improve an outline for a Wikipedia page. You already have a draft outline that covers the general information. Now you want to improve it based on the information learned from an information-seeking conversation to make it more informative.
Here is the format of your writing:
1. Use "#" Title" to indicate section title, "##" Title" to indicate subsection title, "###" Title" to indicate subsubsection title, and so on.
2. Do not include other information.
3. Do not include topic name itself in the outline.

5. Core Code Pattern: Conversation Simulator

class ConvSimulator:
    def forward(self, topic, persona, ground_truth_url, callback_handler):
        dlg_history = []
        for _ in range(self.max_turn):
            # Persona asks question
            user_utterance = self.wiki_writer(
                topic=topic, persona=persona, dialogue_turns=dlg_history
            ).question

            # Check termination
            if "Thank you so much for your help!" in user_utterance:
                break

            # Expert answers with grounded information
            expert_output = self.topic_expert(
                topic=topic, question=user_utterance,
                ground_truth_url=ground_truth_url
            )

            dlg_turn = DialogueTurn(
                agent_utterance=expert_output.answer,
                user_utterance=user_utterance,
                search_queries=expert_output.queries,
                search_results=expert_output.searched_results,
            )
            dlg_history.append(dlg_turn)

6. Concurrent Persona Execution

def _run_conversation(self, conv_simulator, topic, considered_personas):
    conversations = []
    max_workers = min(self.max_thread_num, len(considered_personas))

    with ThreadPoolExecutor(max_workers=max_workers) as executor:
        future_to_persona = {
            executor.submit(conv_simulator, topic, persona, ...): persona
            for persona in considered_personas
        }
        for future in as_completed(future_to_persona):
            persona = future_to_persona[future]
            conv = future.result()
            conversations.append((persona, conv.dlg_history))

    return conversations

Key Parameters

Literature Review Workflow Reference

Extracted from AgentLaboratory.

PhD Student Literature Review Phase

Phase Prompt

Your goal is to perform a literature review for the presented task and add papers to the literature review.
You have access to arXiv and can perform two search operations:
(1) finding many different paper summaries from a search query
(2) getting a single full paper text for an arXiv paper.

Available Commands

SUMMARY — Search for papers

search query here

Returns: List of paper summaries matching the query from arXiv.

FULL_TEXT — Get complete paper

arXiv_paper_ID

Returns: Full text of the specified arXiv paper.

ADD_PAPER — Add paper to review

arXiv_paper_ID PAPER_SUMMARY

Adds the paper with your summary to the official literature review.

Rules

• Single command per turn
• Use ADD_PAPER when you find a relevant paper
• Don't use SUMMARY too many times (be targeted)
• Include three backticks at top and bottom of commands

Workflow Pattern

def literature_review(self):
    arx_eng = ArxivSearch()
    resp = self.phd.inference(self.research_topic, "literature review", step=0)

    for _i in range(max_tries):
        feedback = ""

        if "```SUMMARY" in resp:
            query = extract_prompt(resp, "SUMMARY")
            papers = arx_eng.find_papers_by_str(query, N=num_papers)
            feedback = f"Papers related to '{query}':\n{papers}"

        elif "```FULL_TEXT" in resp:
            paper_id = extract_prompt(resp, "FULL_TEXT")
            full_text = arx_eng.retrieve_full_paper_text(paper_id)
            feedback = full_text  # with expiration marker

        elif "```ADD_PAPER" in resp:
            paper_info = extract_prompt(resp, "ADD_PAPER")
            feedback, text = self.phd.add_review(paper_info, arx_eng)

        # Check if enough papers collected
        if len(self.phd.lit_review) >= num_papers_target:
            lit_review_sum = self.phd.format_review()
            return

        resp = self.phd.inference(
            self.research_topic, "literature review",
            feedback=feedback, step=_i + 1
        )

Review Entry Structure

Each paper in the review contains:

{
    "arxiv_id": "2301.12345",
    "title": "Paper Title",
    "summary": "Your summary of the paper and its relevance to the research topic"
}

Integration with Paper Writing

The formatted literature review (lit_review_sum) is passed to: 1. Plan Formulation — Postdoc uses it to guide experiment planning 2. Paper Writing — Provided as context for Related Work and Introduction sections 3. Result Interpretation — Used to compare findings with prior work

Key Parameters