research-team

A Claude Code plugin that runs research as a pipeline instead of a single prompt. It splits a question into key questions, sends a team of subagents to collect evidence, and has a second set of agents attack that evidence from the sceptic's side. No report ships until deterministic scripts and an auditor agent have both passed it.

Every claim in the final report carries the ID of an evidence unit. Every evidence unit carries a verbatim quote, a full absolute URL, and a source grade. A Python script re-fetches each URL and checks the quote against the page character by character, so a citation that was paraphrased, invented, or attached to the wrong source fails before a human ever reads it.

The report's structure is English; its prose follows the language you asked in. The section headings, the table headers and the confidence labels are always English, so one set of checks reads every report, whatever language the request used. The body of the report, the research brief and the claim text are written in the language of the request, which the evidence ledger records in its top-level deliverable_language field. Verbatim quotes stay in the language of the source. A collection agent picks the language of each query for itself, searching wherever the best sources for its key question happen to be written.

What it does

Five stages.

1. Planning — decompose into 2-7 key questions, state two or more competing hypotheses, and write down what would be observed if each hypothesis were false, before searching. A linter checks the plan and estimates the session's WebSearch budget.
2. Parallel collection — one agent per key question, launched in a single message. research-collector handles the general web (statistics, IR filings, reporting); research-scholar handles academic literature through OpenAlex, arXiv, PubMed and Semantic Scholar. Each agent must clear a floor of queries, independent sources, and disconfirmation queries before it may return.
3. Independent verification — research-verifier takes three claims at a time, in a context that never saw the collection, and tries to break them. Four questions per claim: does the citation exist, does the source actually say this, what does a disconfirming search return, does a second independent source corroborate it.
4. Synthesis — the writer works from per-key-question evidence extracts rather than one long context, keeps fact and interpretation in separate layers, and confines every estimate to a seven-level confidence vocabulary mapped onto probability bands.
5. Audit — deterministic scripts check the ledger and the report, then research-auditor grades both against a 15-item binary rubric and samples atomic facts. A failed critical item sends specific key questions back to collection.

Three modes. LIGHT answers a single factual question in the main session and writes no files. STANDARD runs the pipeline with reduced floors for one decision. DEEP runs the full pipeline with the audit gate. A study escalates from STANDARD to DEEP when sources conflict or the stakes turn out to be serious; it never de-escalates.

Collection floors, per key question:

	DEEP (decision)	DEEP (background) and STANDARD
Search queries	12+	6+
Independent sources	5+	3+
Disconfirmation queries	2+	1+

DEEP holds research-collector and research-scholar to these floors separately. STANDARD counts the total across both roles for one key question, and asks each role for at least three queries and one disconfirmation query. LIGHT sets no query floor.

Source grades. A is governments, international organisations and peer-reviewed papers; B is industry bodies, corporate IR and research firms; C is trade media, reporting and blogs. The pipeline rejects a key figure that rests only on grade-C sources.

Five reference files under plugins/research-team/skills/research-team/references/ state the design rules, and methodology_sources.md holds the academic grounding for them — every citation with a verified DOI or arXiv ID.

Install

/plugin marketplace add tomkd555/research-skill
/plugin install research-team@tomkd555

Then ask for research in the ordinary way, in any language: "research X and give me sources", "look into X", "compare these and help me choose", "summarise it with evidence". The skill picks the mode and launches research-team-lead, which runs the pipeline in its own context.

What you get

A run writes into research/{YYYYMMDD}-{topic-slug}/:

File	Content
research_brief.md	The plan: key questions, competing hypotheses, disconfirmation plan, stopping rules
evidence_log.json	The ledger: every evidence unit, the disconfirmation records, the search log, the gaps
pages/	The cached text of every fetched page, so a quote can be re-checked without a second request
report.md	The report, every claim linked to a ledger ID
audit_result.json	The script findings and the auditor agent's verdict

Requirements

• Claude Code with plugin support.
• Python 3 on PATH. The scripts use the standard library only — no packages to install. Tested on 3.11.
• WebSearch and WebFetch enabled. Claude Code caps WebSearch at 200 calls per session by default and then returns a budget message instead of results, so the planning linter estimates the run's query total and fails a plan that would exceed the cap. Raise CLAUDE_CODE_MAX_WEB_SEARCHES_PER_SESSION or cut key questions.
• Optional: set OPENALEX_MAILTO to a contact address. OpenAlex routes callers who identify themselves to a faster pool. The scripts send the address only when you set the variable, and the paper search works either way.

Layout

.claude-plugin/marketplace.json
plugins/research-team/           everything an install copies
  .claude-plugin/plugin.json
  agents/                        5 agent definitions
  skills/research-team/          SKILL.md, references/, scripts/, assets/
  skills/literature-review/      paper_search.py and its own skill, bundled
tests/                           repository only, never installed

skills/research-team/scripts/labels.py holds one English set of section names, confidence labels and prose patterns. The generator that writes the scaffold, and every check script that reads a deliverable, take their vocabulary from it, so none of them needs a language flag.

literature-review ships here because research-scholar calls its paper_search.py for every academic search. It is a usable skill on its own: a literature review with no web-information side costs a fraction of the full pipeline.

Tests

for t in tests/test_*.py; do python "$t"; done

The tests live outside plugins/, so installing the plugin copies none of them. Ten files cover the check scripts and one covers the paper search. They use the standard library and a local HTTP server; none of them reach the network.

License

MIT. See LICENSE.