Skip to content

Agent Research Skills - Setup Guide

Source: lingzhi227/agent-research-skills via skills.sh - 57.8K combined installs across 31 indexed listings; first seen Feb 20, 2026 GitHub: lingzhi227/agent-research-skills (433 stars; no LICENSE file, GitHub reports no license; pushed 2026-02-27; layout skills/<name>/SKILL.md for 31 skills, with scripts/*.py in 16 of them and references/*.md in 28, plus install.sh, commands/research.md, CHANGELOG.md and two design notes) Category: Research / Academic Writing Quality Tier: 🔵 Community (broad, script-backed and coherent, but the published skills.sh audits are mixed: Snyk Fail on citation-management with a HIGH risk finding, Warn from Snyk and Socket on the flagship research skills, and no LICENSE file in the repo)

Agent Research Skills is a 31-skill pipeline for the academic paper lifecycle, from literature search to conference slides. The author describes it as extracted from 17 GitHub repos studying LLM-agent-driven research automation, and the structure shows it: skills are grouped into phases (discovery and planning, method design, experiments, writing, figures and citations, LaTeX and compilation, review and polish) and cross-link through ## Related Skills sections as upstream or downstream.

This is not a prompt pack. Sixteen skills ship Python scripts with argparse CLIs - literature search against Semantic Scholar, arXiv, OpenAlex and CrossRef, PDF extraction, a paper database, BibTeX harvesting and validation, LaTeX compilation with error-fix loops, statistical summaries, figure templates - and the README states the scripts are stdlib-only where possible. paper-assembly is the orchestrator that runs the phases with checkpointing.

Installation

npx skills add lingzhi227/agent-research-skills -g -a claude-code

The -g flag is not optional in practice: the README states the scripts use ~/.claude/skills/ paths and require a global install.

# Post-install: slash commands, dependency check, script syntax verification
git clone https://github.com/lingzhi227/agent-research-skills.git /tmp/agent-research-skills
/tmp/agent-research-skills/install.sh
rm -rf /tmp/agent-research-skills

# Dependencies the README lists, installed only if you use the skills that need them
pip install PyMuPDF        # self-review, deep-research: PDF parsing
pip install numpy scipy    # data-analysis: statistical tests

Semantic Scholar key, if you want the higher rate limits the README mentions: create ~/keys.md with the line S2_API_Key: your-key-here. Python 3 is required for every script.

What It Provides

Skill Installs Use For
literature-review 3,987 Multi-perspective literature review: generate expert personas, run grounded Q&A dialogues and synthesise the findings into structured knowledge
literature-search 2,608 Multi-source academic search across Semantic Scholar, arXiv, OpenAlex and CrossRef, returning JSONL with title, authors, venue, citations and BibTeX
figure-generation 2,242 Publication-quality matplotlib figures through a three-phase pipeline: query expansion, code generation with execution, then VLM visual feedback
latex-formatting 2,206 Conference templates for ICML, ICLR, NeurIPS, AAAI and ACL, package and formatting fixes, pre-submission checks
math-reasoning 2,146 Derivations, proofs, problem formalisation, statistical test selection and LaTeX notation tables
backward-traceability 2,123 Ties every number in the finished PDF to the code line that produced it, using hypertarget and hyperlink pairs plus num checks
citation-management 2,098 BibTeX harvesting from a draft, cite key validation against the .bib file, deduplication and bibliography formatting

The remaining 24 indexed listings range from 1,519 to 1,965 installs and carry the substance of the pipeline: data-analysis 1,965, deep-research 1,874, idea-generation 1,801, novelty-assessment 1,785, paper-writing-section 1,778, paper-revision 1,771, github-research 1,763, experiment-design 1,762, research-planning 1,744, survey-generation 1,687, algorithm-design 1,680, code-debugging 1,661, rebuttal-writing 1,655, related-work-writing 1,638, table-generation 1,625, self-review 1,616, paper-compilation 1,592, slide-generation 1,591, atomic-decomposition 1,589, paper-to-code 1,572, experiment-code 1,570, paper-assembly 1,568, symbolic-equation 1,546 and excalidraw-skill 1,519. The tight 1,519 to 1,965 band suggests the suite is usually installed as a unit rather than skill by skill.

Why This Matters for Hermes Agents

Research agents fail on provenance. A model asked to summarise a field will produce plausible citations, and a model asked to report numbers will produce numbers with no path back to the run that made them. This suite attacks both: literature-search queries real APIs and returns structured records with BibTeX, citation-management validates cite keys against the actual .bib file, and backward-traceability makes every number in the PDF a hyperlink to the code line that produced it.

The second gain is the review loop. self-review runs the NeurIPS review form with three reviewer personas and ensemble scoring, paper-revision maps reviewer concerns onto sections, and rebuttal-writing produces point-by-point responses. That is the part of paper work agents are usually worst at, because it needs a hostile reader rather than a helpful one, and the personas supply the hostility on demand.

The third is that the scripts make the skills executable rather than advisory. paper-compilation runs pdflatex and bibtex with an auto-fix loop, latex-formatting checks venue compliance before submission, table-generation converts JSON or CSV into booktabs tables with bold-best and significance stars. An agent can run these and read the output, which turns "looks done" into "compiles clean".

The tradeoff is real and it belongs in the same breath: this is the most script-heavy suite in the academic category, and it carries the only Snyk Fail found in this batch. Read the Security section before installing it near a credential store.

Usage

You say What happens
"Do a literature review on protein folding with LLMs" literature-review runs the multi-persona dialogue workflow and literature-search pulls the underlying papers from the four indexes
"Search arXiv, OpenAlex and Semantic Scholar for long-context attention" literature-search queries all configured sources and returns ranked JSONL with citations and BibTeX
"Make publication figures from results.csv" figure-generation writes matplotlib code, executes it and reviews the rendered output through a VLM feedback loop
"Format this draft for NeurIPS submission" latex-formatting applies the venue template and runs the pre-submission checks
"Check whether my idea is actually novel" novelty-assessment runs multi-round search and critique with a harsh critic persona and returns a binary decision with justification
"Write the rebuttal for these reviewer comments" rebuttal-writing maps each concern to an evidence-based response, and paper-revision applies the accepted edits
"Every number in my PDF needs to trace to code" backward-traceability instruments the LaTeX so each reported value hyperlinks to the code line behind it

Verification

Read the flagships before installing anything. All four URLs below returned HTTP 200 when checked with curl on Oct 26, 2026:

curl -sL https://raw.githubusercontent.com/lingzhi227/agent-research-skills/main/README.md
curl -sL https://raw.githubusercontent.com/lingzhi227/agent-research-skills/main/skills/literature-review/SKILL.md
curl -sL https://raw.githubusercontent.com/lingzhi227/agent-research-skills/main/skills/deep-research/SKILL.md
curl -sL https://raw.githubusercontent.com/lingzhi227/agent-research-skills/main/skills/citation-management/SKILL.md

Then run the repo's own dependency check, and confirm the scripts actually execute:

./install.sh --check
python3 ~/.claude/skills/literature-search/scripts/search_arxiv.py --help

install.sh verifies dependencies and script syntax, and it tracks a list of 30 skills - github-research (added in 1.1.0) is not in that list, so a clean report does not cover it.

Security

skills.sh verdicts for sampled skills (verified Oct 26, 2026) - re-check the security pages on skills.sh before production use:

Skill Gen Agent Trust Hub Socket Snyk
literature-review Pass Pass Warn
deep-research Pass Warn Warn
citation-management Warn Warn Fail

These are the only non-Pass verdicts in this batch, and they point at one shared cause. Snyk rates citation-management HIGH with W007 (insecure credential handling): the SKILL.md shows a command that extracts the Semantic Scholar key with grep S2_API_Key /Users/lingzhi/Code/keys.md | cut -d: -f2 and passes it straight in as --api-key, which puts the secret in the model's context and the shell history. Gen Agent Trust Hub files the same pattern under CREDENTIALS_UNSAFE and DATA_EXFILTRATION, with INDIRECT_PROMPT_INJECTION and EXTERNAL_DOWNLOADS alongside; Socket raises a low-severity anomaly for "forwarding a plaintext API key from a personal file into unseen local cross-skill scripts". The second finding, W011, applies to the literature skills generally: they fetch third-party papers chosen by generated queries, so untrusted text enters the context by design.

Beyond the verdicts, the permissions the files ask for are substantial. Sixteen skills ship Python scripts; several of them make outbound calls to Semantic Scholar, arXiv, OpenAlex, CrossRef, GitHub and Papers with Code, and write downloaded PDFs and JSONL databases to disk. paper-compilation shells out to pdflatex and bibtex. excalidraw-skill needs a running canvas server (EXPRESS_SERVER_URL, default http://localhost:3000) and drives it over MCP with element-level writes and deletes. No skill declares an allowed-tools block, so the host's tool policy is the only brake. If you install this suite, create the key file yourself, keep the key out of the prompt (export it as an environment variable and edit the SKILL.md examples), and run the writing and compilation skills in a directory you are willing to have overwritten.

Integration with CorpusIQ

CorpusIQ is the connector layer that lets AI assistants read the systems a business already runs on - accounting, payments, CRM, email, files and analytics - and answer with the source named. It is read-only by design: it retrieves data and never writes to, or acts in, a connected system, which is what makes it safe to point at live accounts.

This page is part of the CorpusIQ directory, the reference we maintain for the MCP server and agent-skill ecosystem. Run alongside CorpusIQ, a third-party server like this one can be combined with a company's own business data in a single answer.

See what CorpusIQ connects

Limitations

  • No LICENSE file, and GitHub reports no license, so the default all-rights-reserved position applies. Ask the author before using the skills or scripts in anything redistributed.
  • Hardcoded personal paths. citation-management and literature-search read /Users/lingzhi/Code/keys.md, and deep-research writes to /Users/lingzhi/Code/deep-research-output/{slug}/. The README documents ~/keys.md and ~/deep-research-output/ instead, so the docs and the skill files disagree; edit both before the first run.
  • The README's design-notes link points at /Users/lingzhi/Code/research-engine/SKILLS_DESIGN.md, which is the author's local filesystem and does not resolve for anyone else.
  • install.sh checks 30 skills while the tree holds 31: github-research, added in 1.1.0, is missing from its list, and its output says "30 skills" either way.
  • Runtime dependencies are uneven. PDF parsing needs PyMuPDF, statistics need numpy and scipy, compilation needs a working TeX distribution, and the diagramming skill needs a separate canvas server. A bare install gets you prompts and no tooling.
  • Audit verdicts are mixed and one is a Snyk Fail with a HIGH risk rating (see Security). The suite is best treated as a research assistant that runs in a sandbox, not as a trusted dependency.
  • Snapshot data, verified Oct 26, 2026: 57,762 combined installs across 31 indexed listings; 433 GitHub stars; no license; last pushed 2026-02-27. Counts drift over time.

Pro Tips

  • Set up the key file before anything else, and patch the two SKILL.md examples that grep /Users/lingzhi/Code/keys.md. Exporting S2_API_KEY in the environment and referencing that variable keeps the secret out of the prompt, which is what the auditors flagged.
  • Install with -g. Half the suite's value is the scripts at ~/.claude/skills/<skill>/scripts/, and a project-local install leaves those paths empty.
  • Point the output directory at your own project. deep-research defaults to an absolute path on the author's machine, so set the output root explicitly on the first run or the files land nowhere you can find.
  • Run literature-search first and check the JSONL before you run literature-review. The review skill reasons over whatever the search returned, and a bad query propagates into a confident review of the wrong literature.
  • Use paper-compilation as the done-check. It runs pdflatex and bibtex with an auto-fix loop, so "it compiles" becomes evidence rather than an opinion.
  • Leave excalidraw-skill out unless you actually need canvas diagrams: it is the only piece that requires a long-running local server.

← Skills Catalog | Skills Marketplace → Powered by CorpusIQ

AI answers on your data. See where every number comes from.

Connect your tools and ask in ChatGPT, Claude, or Perplexity. Enterprise-grade answers in minutes. From $29.95/month.

Start free, no credit card
Contact Terms Privacy Pricing