How to use the Systematic Literature Review agent skill

What does the systematic literature review skill make the agent do?

The skill gives the agent a fixed five-phase workflow for multi-paper surveys. It plans the scope with you, runs one arXiv search through a bundled script, has helper subagents extract key details from each paper (research question, methodology, key findings and limitations), synthesises themes across the set, and writes a final report with consistent citations.

It is built for breadth across many papers, not depth on one. For a single paper, the skill itself routes you to academic-paper-review instead.

• Default paper count is 20, with a hard upper bound of 50.
• Citation formats: APA, IEEE or BibTeX.
• Searches arXiv only.
• Also covers annotated bibliographies and cross-paper comparisons.

How do you install the skill?

Download the skill ZIP from the skill's page in this directory. You can then install it in one of two ways.

For Claude Code, unzip the download into your personal skills folder ~/.claude/skills/, or into a project's .claude/skills/ folder if you want it available only in that project. For the Claude apps, upload the ZIP under Settings, then Capabilities, then Skills.

• Personal use across projects: unzip into ~/.claude/skills/.
• Single project only: unzip into the project's .claude/skills/ folder.
• Claude apps: upload the ZIP under Settings → Capabilities → Skills.

What do you need before you start?

The skill has three requirements. Python must be available so the bundled script scripts/arxiv_search.py can run. The script uses requests if it is installed and falls back to urllib otherwise. The machine also needs internet access to the public arXiv API, which requires no API key. Finally, your host agent must support a subagent task tool, because the extraction phase delegates work to subagents.

• Python available to run scripts/arxiv_search.py.
• Internet access to the public arXiv API (no API key).
• Subagent task tool support in the host agent.

How do you trigger the skill with a natural request?

You do not need a special command. Ask for a survey, review, synthesis or annotated bibliography in plain language, and mention a citation format if you have one. Requests such as 'survey transformer attention variants' or 'do an SLR on this topic in APA format' match the skill's description.

Before searching, the agent asks one clarifying question covering any missing details: topic, number of papers, optional time window, optional arXiv category, citation format and output location. Answer them in one reply. If you give a single paper URL and ask for a review, the agent should route you to academic-paper-review instead.

• 'Survey recent work on diffusion models for images, APA format.'
• 'Make an annotated bibliography on graph neural networks for molecules.'
• 'Compare methodologies across papers on retrieval-augmented generation.'
• 'Review this paper' with one URL will route to academic-paper-review, not this skill.

What does a typical search look like?

The agent reduces your topic to two or three core keywords and passes them to the search script. Qualifiers such as 'in computer vision' belong in the category filter rather than the query, because long exact phrases often return zero results. The script prints a JSON list of papers with title, authors, abstract, dates, categories and links. The skill uses relevance sorting by default. Date sorting returns the newest papers in a category regardless of topic, so it is only used when you explicitly ask for chronological order.

Extraction then runs in batches of about five papers through subagents, at most three at a time. The final report is saved to the default output location /mnt/user-data/outputs/ unless you choose another.

• Good query: "diffusion models" with category cs.CV.
• Poor query: "diffusion models in computer vision" as one exact phrase.

What are the safety notes?

The skill runs a script and uses the network, so review it before you install it. The bundled script calls the public arXiv API. The skill is designed to save only the final report to disk; search results stay in the conversation. If a search fails because of a network error, the agent is meant to report the error and stop rather than invent paper details.

The skill is not a substitute for expert judgement or peer review, and it is not a PRISMA-compliant process. Treat its output as a draft to check against the papers themselves.

• Runs scripts: yes. Uses network: yes.
• Only the final report is written to disk.
• Verify every citation and claim against the original paper before use.

What tips help you get better results?

Keep your topic narrow enough to have a manageable set of papers, and ask for a sub-topic split if you want more than 50. Use a time window when you want recent work, and state the citation format up front so the output matches your venue. Remember that only arXiv is searched, so journal-only or paywalled work will be missing from the review.

• Use 2-3 core keywords in your topic, and put field or date limits in the category or time window.
• For more than 50 papers, split the survey by sub-topic.
• If you need BibTeX for LaTeX, say so; arXiv preprints are cited as @misc entries.
• If the search returns few papers, the agent reports the real count rather than padding the list.