deep-research — Claude Code plugin: multi-phase research
A Claude plugin that runs research through nine phases and demands at least three independent sources per thesis. Built for questions where a wrong answer costs more than the tokens.
/plugin marketplace add alirezarezvani/claude-skills
/plugin install deep-research@claude-code-skillsWhat deep-research does
deep-research sits at the heavyweight end of the research domain. The fast research router classifies a query by keyword and hands back a short brief; this one charges for rigor instead — falsifiable hypotheses, parallel sub-agent fan-out, source triangulation, and a mandatory adversarial pass.
The pipeline has nine phases. It starts by fixing the underlying decision and stating two to four falsifiable hypotheses, then picks a report genre and writes plan.md with scope, sourcing strategy, opposition queries and stop-criteria. From there it audits available API keys and channels, fans out sub-agents to fetch and dedup sources, scores each on credibility, recency and bias, synthesizes with four self-critique questions, runs a lightweight citation check, and emits refresh_targets.md for delta updates. Every source lands in its own file under sources/NN.md with verbatim quotes.
A thesis backed by fewer than three independent, differently-typed sources is labelled insufficient evidence rather than fact. Empty fetch means empty claim, so fabricated citations do not happen. The adversarial pass is required on medium and deep investigations. What you get is a reusable folder, not a wall of chat. For quick fact-checks, 12-dimension competitor scoring or low-risk overviews this is overkill.
What it helps with
- stress-test a strategic hypothesis before committing budget
- compare several options and see which one survives the evidence
- map a field with quotes you can go back and verify
- steel-man the counter-arguments before publishing a claim
- refresh an earlier report by deltas instead of starting over
Example prompts for the agent
- /cs-deep-research test the hypothesis that this stack pays for itself within a year
- Run a deep investigation comparing three architecture options and flag where evidence is thin
- Build a market report with a quote per source and a list of refresh targets
How to install deep-research
Plugin marketplace
/plugin marketplace add alirezarezvani/claude-skills/plugin install deep-research@claude-code-skillsStep by step, with commands for each platform: Claude Code plugins: marketplaces, installing, and building your own
Questions about deep-research
How do I install deep-research in Claude?
Add the marketplace with /plugin marketplace add alirezarezvani/claude-skills, then install the plugin: /plugin install deep-research@claude-code-skills.
What does deep-research do?
It walks a question through nine phases: reframing into falsifiable hypotheses, planning, capability discovery, parallel source collection, scoring and triangulation, synthesis with an adversarial pass, citation verification, and refresh targets. The output is a folder with one file per source and verbatim quotes.
How is deep-research different from the fast research router?
The research router classifies by keyword and returns a short brief for low-risk decisions. deep-research costs more: it requires at least three independent, differently-typed sources per thesis and a mandatory adversarial pass.
Source text
deep-research · MIT
# deep-research — Disciplined Meta-Research > The heavyweight end of the research domain: an auditable, reusable, triangulated investigation — for when a wrong answer costs more than the tokens spent getting it right. `deep-research` is the rigor-first alternative to the fast [`research` router](../research/). Where the router does keyword-classify → delegate → short brief for low-decision-risk questions, `deep-research` pays for rigor: falsifiable hypotheses, parallel sub-agent fan-out, source triangulation, a mandatory adversarial pass, and per-source files with verbatim quotes. ## The 9-phase pipeline | # | Phase | What it does | |---|-------|--------------| | 1 | Reframe | Fix the underlying decision; state 2-4 falsifiable hypotheses | | 2 | Genre & blocks | Pick report genre + building blocks | | 3 | Plan | `plan.md`: scope, sourcing strategy, opposition queries, risk register, stop-criteria | | 3.5 | Capability discovery | Audit available API keys/channels; map subtopics to sources | | 4 | Search (parallel) | Fan out sub-agents → fetch & dedup → save each to `sources/NN.md` | | 5 | Score & triangulate | Rate Credibility/Recency/Bias; require >=3 independent, differently-typed sources per thesis | | 6 | Synthesize + adversarial | Assemble from blocks; 4 self-critique questions; steel-man counter-arguments | | 6.5 | Verify | Lightweight citation check | | 7 | Refresh targets | Emit `refresh_targets.md` for delta-updates | ## Core discipline - **No fabricated citations** — empty fetch = empty claim; every assertion binds to a saved verbatim quote. - **Triangulation mandatory** — a thesis with < 3 independent, differently-typed sources is "insufficient evidence," not fact. - **Adversarial pass required** on medium/deep investigations. - **Persist to files** — output is a reusable folder, not a chat wall. ## Use / don't use **Use for:** strategy, comparing N options, hypothesis validation, mapping a field. **Not for:** quick fact-checks (answer directly), 12-dimension competitor scoring (`competitive-teardown`), or fast low-risk overviews (the research router). ## What's in the box - **Skill:** [`skills/deep-research/SKILL.md`](skills/deep-research/SKILL.md) - **Agent:** [`agents/cs-deep-research.md`](agents/cs-deep-research.md) - **Command:** [`commands/cs-deep-research.md`](commands/cs-deep-research.md) — `/cs-deep-research` - **Reference:** [`skills/deep-research/references/full-catalog.md`](skills/deep-research/references/full-catalog.md) — pointer to the upstream source catalog ## Attribution Methodology contributed by [@Socialpranker](https://github.com/Socialpranker) (PR #851). The continuously-maintained source catalog (29 channels, 460+ statistical sources, 39 validated APIs, 103 report blocks) lives upstream at [claude-deep-research](https://github.com/Socialpranker/claude-deep-research); the methodology here is self-contained. MIT.
© Alireza Rezvani · MIT · alirezarezvani/claude-skills
