Skip to main content
skill CX Codex

deep-research · 5dive-ai — Claude Code skill: structured fact-checking across sources

deep-research runs multi-phase research with source verification and explicit confidence scoring. Use when a single lookup is not enough and getting it wrong costs — option comparison, fact validation, market analysis.

MIT 268 installs ★ 2 updated 2026-10-23

Install into Claude Code · Skills CLI (Vercel) All methods and platforms ↓
npx skills add 5dive-ai/skills --skill deep-research -a claude-code
Open on GitHub

What deep-research does

The skill breaks research into four stages: first it clarifies the question and identifies key sub-questions, then it searches across multiple sources and tests each claim for weaknesses, separates facts from inferences, and closes with a brief showing confidence level and remaining gaps.

It includes checklists for assessing source reliability — an author outside their domain ranks lower than a primary document, sponsored research gets extra scrutiny. The skill marks where each number came from, hunts for the opposing view, and checks the math and units. Default position: treat every key claim as guilty until corroborated, and look for what would break it, not what confirms it.

Use it before decisions where certainty is contested — picking between tools before rollout, validating vendor claims, sizing a market opportunity for an investment call, comparing options under different constraints. Output is a ready brief: opening with the conclusion, evidence backing each point with citations, and an honest map of what remains unknown.

What it helps with

  • collect and verify facts across sources before a major decision
  • tell real data from vendor marketing claims
  • compare multiple options against clear, verified criteria
  • assess source reliability and how current internet information is
  • synthesize a conclusion with explicit confidence and unknowns

Example prompts for the agent

  • Should we pick Next.js, Remix, or Astro for the new React project — what does the real data say?
  • Check the speed claims for this new data parser against Pandas and Polars benchmarks.
  • How large is the no-code platform market, and which vendors have real traction?

How to install deep-research

1Skills CLI (Vercel)

npx skills add 5dive-ai/skills --skill deep-research -a claude-code

2Manually: put the files in place

~/.claude/skills/deep-research/SKILL.md

Copy the whole skill folder; for one project use .claude/skills/deep-research/

Step by step, with commands for each platform: How to install skills in Claude Code: five ways

Questions about deep-research

How do I install deep-research in Claude Code?

Run npx skills add 5dive-ai/skills --skill deep-research -a claude-code, then find the installed skill at ~/.claude/skills/deep-research/SKILL.md.

What does deep-research do?

It runs multi-phase research: clarifies the question, gathers from multiple sources, tests each claim for weaknesses, separates facts from inference, and gives explicit confidence in the answer.

How is this different from a regular web search?

A regular search returns top links. This skill actively hunts for disagreement, checks the math and sources, flags where data may be stale, and marks what is still unknown — the output is ready for a decision, not just informational.

Source text

SKILL.md · MIT
---
name: deep-research
description: >-
  Run structured multi-source research that produces a cited, adversarially
  verified synthesis with explicit confidence. Use when a question needs more
  than a single lookup and being wrong is costly — comparing options,
  validating a claim, surveying a market, briefing a decision.
version: 1.0.0
license: MIT
---

# Deep Research Guide

Covers scoping (and narrowing) the question, fanning out across sources,
judging source quality, adversarially verifying each key claim, separating fact
from inference, and reporting with explicit confidence.

Keywords: deep research, multi-source, fact-check, verify, synthesize, cited
report, source credibility, due diligence.


Produce a **trustworthy synthesized answer**, not a pile of links. The output of research is a
decision-grade brief: a clear claim, the evidence behind it, citations, and an honest map of
what is still unknown. Speed without verification is a liability — a confident wrong answer
costs more than a slow right one.

> Core principle: default to skeptical. Treat every key claim as guilty until corroborated.
> Your job is not to confirm a hypothesis — it's to find what would break it.

## Workflow — plan → gather → verify → synthesize

Run these as distinct phases. Don't start writing the report while you're still gathering, and
don't gather before the question is scoped.

### 1. Plan — scope the question

Before searching, pin down what's actually being asked. Most bad research answers a question
nobody asked.

- **State the question in one sentence.** If you can't, it's not scoped yet.
- **Name the decision it serves.** "Which X should we pick?" needs different evidence than
  "Is X true?". The decision sets the bar for confidence.
- **List the answer's shape.** A number? A recommendation? A comparison table? A yes/no with
  caveats? Knowing the shape tells you when you're done.
- **Set boundaries**: time window (recency that matters), geography, scale, definitions of
  fuzzy terms.

**Narrow an underspecified question before spending effort.** If the ask is "what should I
buy / which tool / is this a good idea" without budget, use-case, constraints, or context, ask
2–3 clarifying questions first. Researching the wrong question thoroughly is still wrong.

Then decompose into **sub-questions** — the 3–7 things that must each be answered for the whole
to hold. Research the sub-questions; assemble the answer.

### 2. Gather — fan out across sources

- **Cast wide before going deep.** Run several differently-worded queries; don't anchor on the
  first source's framing. Search for the *counter*-claim too ("X is overrated", "problems with
  X") — not just confirmation.
- **Go to primary sources.** Prefer the original study, filing, spec, dataset, or official
  doc over an article summarizing it. Summaries drift; numbers get garbled in retelling.
- **Triangulate.** A claim is only as strong as the number of *independent* sources that
  confirm it. Three outlets all citing one press release is one source, not three.
- **Capture as you go**: for each fact, note the source, the date, and a direct quote/figure.
  You cannot cite what you didn't record.

### 3. Verify — adversarial check (the step people skip)

For each **key claim** (the ones the conclusion rests on), actively try to refute it:

- **Find the origin.** Trace the claim to its source. Where did this number actually come
  from? Who measured it, how, and when?
- **Look for the strongest disagreement.** Who says the opposite, and why? A claim you can't
  find any dissent on is either settled or you haven't looked hard enough.
- **Check the math and the units.** Percentages without a base, totals that don't add up,
  growth rates with no time frame, and apples-to-oranges comparisons are the common tells.
- **Test recency.** Is this still true? Prices, rankings, "fastest/largest/only" claims, and
  policy facts decay. A correct 2019 fact can be a wrong 2026 answer.
- **Watch for self-interest.** Vendor benchmarks, sponsored studies, and anything selling
  something get an extra round of scrutiny.

If a claim survives a genuine attempt to break it, it's load-bearing. If it doesn't, demote it
to "reported but unverified" or drop it.

### 4. Synthesize — write the brief

- **Lead with the answer.** First line: the conclusion / recommendation. Decision-makers
  read top-down and may stop after the first paragraph — make it count.
- **Then the why**, structured by sub-question, each point carrying its citation.
- **Separate fact from inference explicitly** (see below).
- **Close with confidence + unknowns.**

## Source-credibility checklist

Score each source before you lean on it:

- **Primary or secondary?** Original data/document > reporting on it > commentary on the
  reporting.
- **Authority** — does the author/org actually have standing on this topic, or are they out of
  their lane?
- **Recency** — is it current enough for a claim that changes over time? Note the date, always.
- **Independence** — funded by, owned by, or selling the thing in question? Conflicts bias.
- **Method transparency** — can you see how they got the number (sample, methodology, sources),
  or are you trusting an assertion?
- **Corroboration** — do independent sources agree? Outliers need explaining, not silent
  dropping.
- **Track record** — has this source been reliable/retracted before?

Rough hierarchy (context-dependent, not absolute): peer-reviewed studies, official
filings/standards, and primary datasets at the top; reputable journalism and expert analysis
in the middle; anonymous posts, marketing, and AI-generated content summaries near the bottom.
A low-tier source can still be right — it just needs corroboration before it carries weight.

## Separating fact from inference

Be ruthless about which is which; conflating them is how research misleads.

- **Fact** — directly stated by a credible source, with a citation. ("Revenue was $4.2M in
  2025 [source].")
- **Inference** — your reasoning *from* facts. Label it. ("This implies ~30% YoY growth,
  assuming the 2024 figure of $3.2M is comparable.")
- **Assumption** — something you're taking as given without evidence. Name it so the reader can
  challenge it. ("Assuming the same accounting basis across years.")
- **Unknown** — a gap you couldn't fill. State it; don't paper over it.

Phrases that signal you're doing it right: "according to…", "this suggests…", "I could not
find…", "sources disagree on…".

## Reporting confidence and unknowns

End every brief with an explicit confidence statement. Vague hedging ("seems like") is useless;
calibrated confidence is actionable.

- **High** — multiple independent primary sources agree; recent; verified the underlying math.
- **Medium** — corroborated but with gaps, dated data, or some reliance on secondary sources.
- **Low** — single source, conflicting evidence, stale data, or heavy inference. Say so loudly.

Always include a short **"What I couldn't verify / what would change this answer"** section.
Naming the unknowns is a feature: it tells the decision-maker where the risk lives and what to
check before betting on it.

## Report template

```
ANSWER:        <the conclusion / recommendation, one or two sentences>
CONFIDENCE:    High / Medium / Low — <one-line why>

KEY FINDINGS
  1. <claim> [source, date]
  2. <claim> [source, date]   (note: sources disagree — see below)
  ...

REASONING / INFERENCE
  <what you concluded from the facts, with assumptions named>

CONTEXT & CAVEATS
  <scope, definitions, anything that bounds the answer>

UNKNOWNS / WHAT WOULD CHANGE THIS
  <gaps you couldn't fill; what to verify before acting>

SOURCES
  [1] <title> — <publisher/author>, <date>, <url> — primary/secondary, why trusted
  ...
```

## Anti-patterns

- **Confirmation hunting** — searching only for what you hope is true. Search the opposite.
- **Citation laundering** — three articles citing one origin presented as three sources.
- **Stale-fact trap** — quoting a "current" superlative that quietly expired.
- **Burying the answer** — making the reader mine paragraphs for the conclusion.
- **False precision** — "$4,231,847" from a source that said "about $4M".
- **Unlabeled inference** — presenting your reasoning as if it were a sourced fact.
- **Over-hedging** — refusing to give an answer when one is warranted. Calibrate, don't dodge.

© 5dive-ai · MIT · 5dive-ai/skills

Similar skills

All skills for Agent setup: memory, context, orchestration →

deep-research by other authors comes first.

deep-research Imbad0202 deep-research runs a team of 13 agents through academic research on any topic, from framing the question to an APA 7.0 report. CX Codex +1 8,200installs deep-research affaan-m deep-research — a skill for Claude Code, Codex, Cursor and other agents that builds cited research reports from 15–30 sources using firecrawl and exa MCP tools. Triggers on phrases like "research", "deep dive", or "competitive analysis". CX Codex +1 7,000installs deep-research bytedance deep-research — a skill for Claude Code runs source-based research with current evidence and cross-checks. It fits questions that need more than one search. CX Codex +1 2,800installs deep-research 199-biotechnologies deep-research — a skill for Claude Code, Codex, Cursor and other agents that runs a multi-phase research pipeline and produces a fully cited report with per-claim verification. CX Codex +1 10,500installs deep-research ruvnet deep-research orchestrates multi-phase research using web search, codebase analysis, and memory storage. It generates structured findings with evidence ratings for complex topics. CX Codex +1 894installs research-deep Eugeniu Ghelbur research-deep scans your Obsidian vault to map what you already know, identifies missing or stale information, fills gaps via Perplexity or Grok, synthesizes the findings, and updates your notes on people, projects, and ideas automatically. CX Codex +1 ★ 5,412repo stars multi-source-investigation poemswe multi-source-investigation — a skill that searches the web, academic databases, and local data to verify claims across independent sources. Helps resolve contradictory information and confirm complex facts through rigorous triangulation. CX Codex +1 147installs research-synthesizer davila7 research-synthesizer — a subagent that consolidates output from multiple specialist researchers into one unified analysis. It detects patterns, surfaces contradictions, and structures complex data to make it clearer and more actionable. CX Codex 1,377installs fact-check jwynia fact-check is a Claude Code skill that checks finished text by extracting claims and comparing them with sources to catch hallucinations. CX Codex +1 799installs tooluniverse-literature-deep-research mims-harvard tooluniverse-literature-deep-research — a skill for deep literature search across PubMed, EuropePMC and bioRxiv, with citation networks and evidence grading. Useful when you need a review with citations. CX Codex +1 677installs researcher bitwize-music-studio researcher is a skill for investigative work that finds primary sources, cross-checks facts across three or more independent sources, and builds evidence chains with quotes and citations. CX Codex +1 253installs

From the same set4

Project sponsors

Companies that keep this analytics open