deep-research · 5dive-ai — скилл для Claude Code: структурированный поиск фактов
Скилл deep-research для Claude Code проводит многоэтапное исследование с проверкой источников и оценкой уверенности. Сравнивает варианты, проверяет факты и анализирует рынок при высокой цене ошибки.
MIT 268 установок ★ 2 обновлён 2026-10-23
npx skills add 5dive-ai/skills --skill deep-research -a claude-codeЧто делает deep-research
Скилл разбирает исследование на четыре фазы: сначала уточняет вопрос и выделяет ключевые подвопросы, потом ищет по разным источникам и проверяет каждое утверждение на слабость, отделяет факты от выводов и завершает краткой сводкой с явной уверенностью.
Внутри встроены чеклисты для оценки надёжности источника — автор статьи вне своей области стоит ниже, чем первичный документ, спонсируемые исследования требуют особой проверки. Скилл помечает, откуда взялась каждая цифра, ищет противоположное мнение и проверяет математику и единицы измерения.
Пригодится для подготовки решений, когда уверенность существует, но оспаривается — например, сравнение инструментов перед внедрением, проверка обещаний продавца, анализ ёмкости рынка для инвестиционного решения. Результат — не список ссылок, а готовая сводка с выводом в начале, ссылками и честной картой неизвестного.
Какие задачи решает
- собрать и проверить факты из разных источников перед ключевым решением
- отделить реальные данные от маркетинговых обещаний продавца
- сравнить несколько вариантов по чётким, проверенным критериям
- оценить надёжность и актуальность информации из интернета
- синтезировать вывод с явной оценкой неопределённости
Примеры запросов агенту
- Какой фреймворк выбрать для нового проекта на React — Next.js, Remix или Astro?
- Проверь заявления про скорость этого нового парсера данных в сравнении с Pandas и Polars.
- Насколько большой рынок низкокодовых платформ в России, и какие игроки там лидируют?
Как установить deep-research
1Через Skills CLI (Vercel)
npx skills add 5dive-ai/skills --skill deep-research -a claude-code2Вручную: положить файлы
~/.claude/skills/deep-research/SKILL.mdПапка скилла целиком; для одного проекта — .claude/skills/deep-research/
1Через Skills CLI (Vercel)
npx skills add 5dive-ai/skills --skill deep-research -a codex2Вручную: положить файлы
.agents/skills/deep-research/SKILL.mdДля всех проектов — ~/.agents/skills/
1Через Skills CLI (Vercel)
npx skills add 5dive-ai/skills --skill deep-research -a cursor2Вручную: положить файлы
.cursor/skills/deep-research/SKILL.mdCursor читает и .agents/skills/, и .claude/skills/
1GitHub CLI
gh skill install 5dive-ai/skills deep-research2Вручную: положить файлы
.github/skills/deep-research/SKILL.md1Через Skills CLI (Vercel)
npx skills add 5dive-ai/skills --skill deep-research -a gemini-cli2Командой платформы
gemini skills install https://github.com/5dive-ai/skills --path deep-researchВручную: положить файлы
.devin/skills/deep-research/SKILL.mdСтарый путь .windsurf/skills/ тоже читается
Через Skills CLI (Vercel)
npx skills add 5dive-ai/skills --skill deep-research -a opencodeClaude.ai
Упакуйте папку в zip и загрузите в Settings → Capabilities → Skills
По шагам, с командами для каждой платформы: Как установить скилл в Claude Code: пять способов
Вопросы о deep-research
Как установить deep-research в Claude Code?
Выполни npx skills add 5dive-ai/skills --skill deep-research -a claude-code, потом найди установленный скилл по пути ~/.claude/skills/deep-research/SKILL.md.
Что делает deep-research?
Проводит многоэтапное исследование: уточняет вопрос, ищет по разным источникам, проверяет каждое утверждение на слабость, отделяет факты от выводов и даёт явную оценку уверенности в ответе.
Чем этот скилл отличается от обычного поиска в интернете?
Обычный поиск берёт первые ссылки. Этот скилл активно ищет противоположное мнение, проверяет математику и источники, раскрывает, где данные могут быть устаревшими, и отмечает, что осталось неизвестным — результат готов для решения, а не просто информационен.
Исходный текст
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
Похожие скиллы
Все скиллы по теме «Настройка агента: память, контекст, субагенты» →Сначала — deep-research от других авторов.