Research Skills: Ten You Can Actually Install

The list first: ten rows, installs below

How to install (commands copied from each project's own page)

# 1 — 59 directories in that repo. Copy the two or three you want, not all of it.
/plugin marketplace add daymade/claude-code-skills
git clone --depth 1 https://github.com/daymade/claude-code-skills /tmp/dcs
cp -r /tmp/dcs/deep-research /tmp/dcs/competitors-analysis /tmp/dcs/fact-checker ~/.claude/skills/

# 2 — note: also named deep-research, same target directory as 1. Pick one.
git clone https://github.com/199-biotechnologies/claude-deep-research-skill.git ~/.claude/skills/deep-research

# 3 — clone the repo first, then copy. English and Chinese sets, plus Codex / OpenCode variants.
cp -r skills/research-en/* ~/.claude/skills/
cp agents/web-search-agent.md ~/.claude/agents/
cp -r agents/web-search-modules ~/.claude/agents/
pip install pyyaml

# 4 — pulls in 31 skills at once. Worth it only if you write papers.
npx skills add lingzhi227/agent-research-skills -g -a claude-code

# 5 — official, and a plugin of its own: it is not inside example-skills.
/plugin marketplace add anthropics/skills
/plugin install discernment-nudge@anthropic-agent-skills

# 6 — one project, two forms. This is the skill install; MCP install is in the repo.
uv tool install paper-search-mcp
mkdir -p ~/.claude/skills/paper-search
curl -fsSL https://raw.githubusercontent.com/openags/paper-search-mcp/main/claude-code/SKILL.md \
  -o ~/.claude/skills/paper-search/SKILL.md

# 7-10 — the retrieval layer. Pick one, at most two.
uvx arxiv-mcp-server                                   # 7 into mcpServers; papers land in ~/.arxiv-mcp-server/papers
claude plugin install exa@claude-plugins-official      # 8 two tools by default
env FIRECRAWL_API_KEY=fc-YOUR_API_KEY npx -y firecrawl-mcp   # 9 twenty-five tools by default
npx -y tavily-mcp@latest                               # 10 four tools, API key required

# Do this immediately after installing
/skill-doctor    # context cost and invocation frequency per skill

Item by item

• 1. The daymade trio (1.4k stars per the GitHub repo page, 2026-09, MIT): the most carefully built evidence chain on this list. The deep-research SKILL.md pins the procedure to phases P0-P7, and the hinge is P3, where it builds a citation registry — deduplicating, numbering, and recording every dropped source with the reason it was dropped. Once the registry is final, the draft may cite only from it. Three hard rules follow: reopen the original for every load-bearing claim; never invent a URL, every locator comes from an actual retrieval; mark [Unverified] or omit rather than speculate. In the same repo, competitors-analysis demands that repository-level findings come from locally cloned code while market-level claims cite a source, and fact-checker only proposes corrections for your approval. The catch: 59 directories, so installing the lot means 59 descriptions in context every turn.
• 2. 199-biotechnologies/claude-deep-research-skill (1.0k stars per the GitHub repo page, 2026-09, MIT): an eight-phase pipeline with a loop-back — Scope, Plan, Retrieve, Triangulate, outline refinement, Synthesize, Critique (which can jump back), Refine, Package. What makes it singular is two check scripts that actually run: validate_report.py (nine checks) and verify_citations.py (DOI / URL / hallucination detection), inside a validate-fix-retry loop capped at three cycles, with a floor of "10+ sources, 3+ per major claim". Multi-provider search needs search-cli and your own key for something like Brave.
• 3. Weizhena/Deep-Research-skills (2.2k stars per the GitHub repo page, 2026-09, MIT): the only one that makes the human gate its first phase. It produces an outline only — which items to investigate, which fields to collect for each — and waits for you to add, cut and approve before phase two searches item by item and consolidates into one markdown report. The repo carries research-en and research-zh plus a Codex build, so portability across harnesses is its selling point. The price is a manual cp install, an extra web-search-agent to copy, and pip install pyyaml.
• 4. lingzhi227/agent-research-skills (350 stars per the GitHub repo page, 2026-09): 31 skills spanning the whole academic lifecycle — literature-search and novelty-assessment, then experiment-design, paper-writing-section, related-work-writing, then self-review, rebuttal-writing, slide-generation. Discovery runs against four APIs: Semantic Scholar, arXiv, OpenAlex and CrossRef. It is the list's example of "the narrower the domain, the more sense a bulk install makes": a pipeline if you write papers, 31 wasted descriptions if you do not.
• 5. discernment-nudge: from the official anthropics/skills (177k stars on that repo per its GitHub page, 2026-09) — the lightest item here and the most underrated. It does no research; it changes the posture in which you receive an answer. After a substantive answer you might act on, it appends two or three follow-up questions tied to specifics in what was just produced, to help you check key facts, probe assumptions and notice missing context — at most once per conversation. Most of its description is spent on when not to fire: trivial how-tos, purely educational explanations, formatting or assembly work, code you will run, creative writing, and anything you already asked it to double-check. That "when not to" clause being longer than the rest is itself a model for writing skills.
• 6. openags/paper-search-mcp (2.7k stars per the GitHub repo page, 2026-09, MIT): one project in two forms — an MCP server, or a Claude Code skill called paper-search, and the skill form needs no MCP configuration at all. It covers more than twenty sources including arXiv, PubMed, bioRxiv, medRxiv, Semantic Scholar, OpenAlex, CrossRef, Europe PMC, dblp, Zenodo, HAL and SSRN, on a free-first strategy: most sources need no key, while Semantic Scholar, CORE and Unpaywall are steadier with one.
• 7. blazickjp/arxiv-mcp-server (3.2k stars per the GitHub repo page, 2026-09, Apache-2.0): 19 tools, running locally. Its value is that the trail is physical — papers land in ~/.arxiv-mcp-server/papers and stay there, you can read a named section of the original LaTeX instead of a summary of a summary, BibTeX comes straight out of arXiv metadata, and topic watches are built in. The core needs no external search service; a Semantic Scholar key only makes citation lookups more reliable.
• 8. exa-labs/exa-mcp-server (5.0k stars per the GitHub repo page, 2026-09, MIT): exposes only web_search_exa and web_fetch_exa by default, with agent_run and others behind an explicit switch. The hosted endpoint works anonymously under rate limits; a key or OAuth lifts them. On the tool-slot budget it is the most restrained item here — compare it with item 9.
• 9. firecrawl/firecrawl-mcp-server (7.5k stars per the GitHub repo page, 2026-09, MIT; the org was formerly mendableai): the most starred and the most sprawling — 25 tools by default: scrape, map, crawl and search, a set of firecrawl_research_* paper tools, plus eight firecrawl_monitor_* ones. Those eight are where the competitive-intelligence thread finally lands on real tooling: create a watch, run a check, see which page of a site changed. The cloud API needs a key; FIRECRAWL_API_URL points it at a self-hosted instance instead.
• 10. tavily-ai/tavily-mcp (2.4k stars per the GitHub repo page, 2026-09, MIT): four tools — tavily-search, tavily-extract, tavily-map, tavily-crawl. The tool count sits between items 8 and 9, which makes it the middle option when you need whole-site traversal but not twenty-five tool definitions. A key is required.

Why these ten do not compare: three layers, not one row

One common install is simply wrong: items 1 and 2 together. Both target ~/.claude/skills/deep-research, so the names collide; and even after you rename a directory, the two descriptions compete for the same trigger ("research this topic", "write a report on"), so which procedure runs is a coin flip. Pick exactly one process-layer skill. That one is not negotiable.

The reverse is more common still: a carefully chosen process layer and no retrieval at all. Item 1's rule about reopening originals for load-bearing claims then lands on default web search, which surfaces the same aggregator pages as before. No procedure is stricter than the eyes it calls. Papers mean items 6 and 7; watching a competitor's site means item 9's firecrawl_monitor_*; wanting accurate search without burning tool slots means item 8.

Item 4 is a third kind of thing: not general research but a disciplinary pipeline. Pointed at market research it feels wrong, because rebuttal-writing and latex-formatting will never fire in that context while still costing description budget every turn. The test is easy: does your output carry a bibliography? If yes, install it; if no, do not.

Three-minute check: six things to look for in SKILL.md

1. Is there a citation ledger separate from the prose? Item 1 builds one at P3 — numbered, deduplicated, and recording dropped sources with the reason they were dropped — and the draft may cite only from it once frozen. Without that layer, citations are decoration on the prose: you cannot audit what it used or what it threw away.
2. Does uncertainty have an exit? Look for [Unverified], confidence markers, an explicit "say you could not find it". A research skill that is not allowed to output "I do not know" will encode uncertainty as certainty, because you left it only one exit.
3. Is there a dating convention? Item 1 fixes AS_OF at P0 and requires a source date on every time-sensitive claim. Without it, last year's figure and this year's get written into the same sentence as equals.
4. Is there a check a machine can run? Item 2's verify_citations.py tests whether a DOI or URL exists; validate_report.py runs nine checks. This is the only thing in the category that catches errors without reading the content. Mind its boundary, though — see section 04.
5. Where can a human intervene? Item 3 makes outline approval a blocking first phase. The most expensive research error is direction, and direction is only cheap to fix in the first few minutes. A skill that involves you only at "please review the report" has no gate at all.
6. What tools does it assume you have? The easiest one to miss. Read what retrieval the SKILL.md refers to: bundled scripts (item 4's four academic APIs, item 2's search-cli), or an MCP server it assumes you already installed? When that assumption fails it does not error — it degrades quietly.

If three of these six have no answer, what you have bought is the layout of research, not research.

The honest costs: where it is expensive, and where it fools you

Money

Anthropic published its own numbers: per its engineering post "How we built our multi-agent research system", 2025-06, Research runs an orchestrator-worker pattern — a lead agent plans, spawns subagents that work in parallel with their own context windows, and merges their findings with a separate citation pass. The return was a 90.2% improvement on their internal research eval for an Opus 4 lead with Sonnet 4 subagents; the price is roughly 15x the tokens of an ordinary chat interaction, and the same post is explicit that this architecture is a poor fit for tightly interdependent work such as coding. Items 1 and 2, with their lead-plus-subagents structure, spend exactly that. Budget a deep report like a compile, not like a question.

Tool slots

Item 9 ships 25 tool definitions by default; item 8 ships 2. Those definitions sit in context every turn whether or not you use them. Count tools before installing a second retrieval server — and the same arithmetic applies to item 1's 59 directories and item 4's 31 skills, where a bulk install is a description tax paid each turn. Run /skill-doctor afterwards: it reports the context cost and invocation frequency of every skill you have.

Where it fools you

The second trap is that format reads as more credible than content. Numbered citations, confidence markers and a "counter-evidence" section are all visual signals of rigor. A skill that only applies the layout, without ever reopening an original, produces something indistinguishable on screen from real work. The six criteria in section 03 are, at bottom, one question asked six ways: is there a ledger behind this layout?

Third, nearly all of these default to the open web, so second-hand aggregators get treated as sources — Skill 21 on this site takes that apart on its own.