Stop ad-hoc Googling. Start documented investigation.
Docs · Install · How it works · Contribute
You: investigate the trade-offs between Postgres logical replication and CDC tooling
Claude: ✓ Reframed your question (3 hypotheses)
✓ Picked genre: decision (comparison + validation)
✓ Wrote plan.md (17 sections)
✓ Checked your env: 4 APIs available, 2 fallback to HTML
✓ Launched 4 sub-agents across 12 channels
✓ Saved 23 sources to sources/ with quotes
✓ Ran adversarial pass (3 counter-arguments)
✓ Report ready: research/postgres-replication-vs-cdc/2026-05-21_decision.md
A Claude Code skill that turns "research this topic" into a 11-phase pipeline with hypothesis testing, parallel sub-agent search, source triangulation, and adversarial review.
The output is a folder you can return to in a month. Every claim traces to a specific source file. The plan documents why you made every choice. No re-research needed.
New here? Start with the Quickstart — install → invoke → first result in ~5 min.
| Without this | With this |
|---|---|
|
One-shot prompt → wall of text Sources lost in chat history No way to detect bias No reuse next time Generic Google results Sources include... (vague) |
17-section Each source = file with verbatim quotes Mandatory adversarial pass + opposition queries Atomic theses in Every claim → |
For Claude Code (CLI)
git clone https://github.com/Socialpranker/deepdive.git \
~/.claude/skills/deepdiveThat's it. Now type any of these in a Claude Code session:
- "Investigate X"
- "Изучи тему"
- "Validate this hypothesis"
For Claude Desktop (Skills enabled)
# Clone
git clone https://github.com/Socialpranker/deepdive.git
cd deepdive
# Package as .skill bundle
zip -r ../deepdive.skill . -x ".*" -x "*.zip"
# Upload via Claude.app → Settings → Skills → Add SkillFor other LLMs (Codex, Gemini, local)
The 11-phase methodology is portable. Load SKILL.md + references/*.md into the LLM's context manually. Skip the sub-agent parts and use separate chat sessions per subtopic.
The skill runs 11 phases in order:
| Phase | Name | What happens |
|---|
| 1 | Reframing | opus / high | | 2 | Genre & block selection | sonnet / medium | | 3 | Plan | opus / medium | | 3.5 | Capability Discovery | sonnet / low | | 3.7 | Plan-review gate | sonnet / low | | 4 | Search | sonnet / medium | | 5 | Claims-ledger + triangulation | haiku / low | | 5.5 | Evidence filter | sonnet / low | | 6 | Synthesis + multi-angle red team | opus / high | | 6.5 | Verify | haiku / low | | 7 | Refresh targets | sonnet / medium |
Each phase runs on a model matched to its task — Opus where reasoning multiplies (1/3/6), Haiku for the parallel fan-out (4). The skill announces the routing and an estimated cost up front, once.
Every phase is transparent: you see what's happening, you confirm key decisions, and you get a folder you can return to. Before any search fires, the plan-review gate (3.7) shows you the reframing, hypotheses, genre, and channels and lets you approve or edit them — strictness scales with mode (deep waits for an explicit go-ahead, medium is a soft check, shallow skips it). Editing the plan before execution is the single highest-leverage step in the whole pipeline — Gemini Deep Research calls plan review its "biggest lever over output quality," and a wrong plan executed perfectly still produces a wrong report.
Reframing (1) doesn't just restate the question — a router classifies its profile (factual / multi-step / relational / comparative / landscape) and that classification picks the decomposition method: factual questions get flat independent subquestions, multi-step ones ("X given Y") get least-to-most leveling, comparative ones get a shared axis matrix with mandatory opposition queries per candidate. Picking the wrong decomposition for a question's shape is a silent failure mode — the router makes the choice explicit instead of defaulting to "flat parallel" for everything.
Phase 4 (Search) isn't a single pass — it's a bounded loop with three cheap safeguards so it doesn't quietly waste budget or silently give up:
- Cheap goal-check — after each round, a Haiku pass tags every subquestion
met/partial/unmetwith a one-line reason. This is what the expensive Opus evaluation reads instead of re-deriving the gap from scratch, and it's what targets the next round's dispatch. - No-progress circuit breaker — two consecutive rounds that add nothing new to the source pool stop the loop immediately, regardless of remaining budget. The unresolved thread goes to Open Questions instead of burning tokens chasing a dead end.
- Least-to-most decomposition — for layered questions ("X given Y"), subquestions are leveled
L1 → L2instead of dispatched flat in parallel: L1 rounds run first, concrete facts they surface get carried forward, and L2 queries are launched already sharpened by that context. Independent subquestions still run flat.
Scoring (5) doesn't stop at the usual Credibility/Recency/Bias rating — it also flags input-level skepticism: a source that measures its own product, self-reports a benchmark, or is directly disputed by another collected source gets a strict caveat: marker (vendor / self-reported / disputed:sNN) before the claim reaches claims.csv, not after synthesis has already built on it. A claim whose key number carries that marker is capped at confidence: medium (or low for an unresolved dispute) — the same rule shape as primary-first sourcing. Vendor benchmarks are the numbers that most often get quietly repeated as fact; catching them on the way in, not in the red team pass at the end, is the point.
For medium/deep depth, the pipeline runs two more machine-checked passes most one-shot research skips entirely:
- Evidence filter (5.5) — a CRAG-style relevance classifier runs on every (claim, source) pair before synthesis and keeps only the quotes that actually support that specific claim. Dumping every found source into synthesis measurably hurts quality (Search-o1 dropped 33%→24% doing exactly that); this is the fix, not a nice-to-have.
- Faithfulness verification (6.5) — beyond checking that a cited link is alive, the skill checks that the source entails the claim it's attached to (RAGAS/ALCE-style claim⊨quote), and writes
SUPPORTED/PARTIAL/UNSUPPORTEDverdicts to.verify/faithfulness.json. Citation fabrication is common enough industry-wide — the Tow Center found a >60% error rate in AI-generated citations — that checking for it, not just for dead links, is a real differentiator.
None of this is enforced by discipline alone: scripts/validate_phases.py reads a finished run's mode: and checks that every phase mandatory for that mode actually left its file artifact (plan.md, claims.csv, evidence/, .verify/*.json, the dated report, ...). A skipped phase fails the check instead of silently passing — the model can't just claim "done." As of finish-up, this check is a blocker, not a suggestion: the skill won't report a research as done on a red gate, symmetrically to how a report isn't "done" without its verification header. sources.csv itself is now built the same deterministic way — scripts/build_sources_csv.py generates it from sources/NN.md frontmatter (with a --check mode for CI) instead of being assembled by hand each run.
Want to compare models head-to-head? The eval harness scores any run on 6 axes.
|
10 categories: FRAME · EXPLAIN · COMPARE · MAP · VALIDATE · ANALYZE · CLOSE · PEOPLE · NUMBERS · CONTEXT Each block has its own template, anti-patterns, and composition rules. |
Named strategies with query patterns + paywall fallbacks:
|
14 cross-industry + 19 industry categories. Each entry: URL · Type · Access · Quality · Limitations · Combine-with · Fallback. Categories: |
||||||||||||||
|
Free no-auth APIs prioritized:
Auth via env vars only — skill never asks for keys inline. |
GitHub Actions cron validates all endpoints + discovers upstream additions:
|
||||||||||||||
|
Per-phase model selection — quality where it multiplies, cheap where it parallelizes:
~$2 instead of ~$8 on a deep run, and higher quality on critical phases. Override with |
Compare research quality across models. Same question, different configs, scored on 6 axes:
Weighted sum with a citation floor — hallucinated sources can't win on depth. Verdict = quality per dollar. |
Ignores env proxies ( Verification runs two layers: liveness (does the source exist) and faithfulness (does it actually entail the claim it's cited for). Faithfulness verdicts — |
A skipped phase fails the check ( Its own inputs are machine-built too: |
Sample output for a typical decision-genre research:
research/<topic-slug>/
├── plan.md # 17-section plan
├── sources.csv # Index with C/R/B scoring
├── sources/ # One file per source
│ ├── 01_vendor-docs.md # Primary, total=14
│ ├── 02_benchmark-paper.md # Academic, total=12
│ ├── 03_industry-report.md # Industry, total=13
│ ├── 04_forum-thread.md # Forum, total=9 (opposition)
│ └── ... (19 more)
├── findings/
│ ├── F1_<atomic-thesis>.md # confidence: high
│ └── F2_<atomic-thesis>.md # confidence: medium
└── 2026-05-21_decision.md # Final report
Final report structure (assembled from the blocks chosen in plan.md):
## TL;DR
- Claim A holds under condition X [confidence: high]
- Claim B holds conditionally on threshold Y [confidence: medium]
- Claim C is disputed by opposition sources [confidence: low]
## Mental model
[How the underlying mechanism works...]
## Falsification criteria
What would disprove H1, H2, H3...
## Verdict conditional
Recommendation IF: <conditions met>
Different recommendation OTHERWISE: <conditions broken>
## Counter-arguments (steel-man)
CA1: "<the strongest opposing claim>" [source: s09]
→ Our answer: <conditions under which CA1 fails>
CA2: ...Every claim is clickable to its source. A month later, you don't re-research — you read.
The catalog is most valuable when it grows. Easy contributions:
| Time | Type | Example |
|---|---|---|
| 15 min | Add a stat source | Add SimilarWeb Pro to consumer_digital |
| 15 min | Improve a query pattern | Better arxiv channel queries for biology |
| 30 min | New search channel | Add patent-search with USPTO+EPO fallback |
| 1-2h | New industry category | Add industries/aerospace.md |
| 2-4h | New report block | Add decision-tree to compare.md |
| Half-day | LLM adapter | Add codex/ folder with adapted protocols |
How is this different from ChatGPT Deep Research / Perplexity?
Those are products — closed UI, fixed flow, opaque source selection. This is open methodology — you control every step, the protocol is markdown you can fork, the source catalog is yours to extend.
They also don't separate sources into files, don't do explicit triangulation, don't run adversarial passes, and don't produce reusable atomic theses. Nor do they filter evidence for relevance before synthesis (feeding a model everything you found measurably hurts quality — Search-o1 dropped from 33% to 24% accuracy doing that) or verify that a cited source actually supports the claim it's attached to, rather than just existing (faithfulness, not just liveness). Citation fabrication is common enough industry-wide — the Tow Center found a >60% error rate in AI-generated citations — that checking for it is a real differentiator, not a nice-to-have.
Honesty about sources goes further than checking they exist: scoring flags a source that's measuring its own product, self-reporting a benchmark, or directly disputed by another collected source, and caps the confidence of any claim resting on that number — before it ever reaches the report. Vendor benchmarks getting quietly repeated as fact is a market-wide problem; catching it on input, not as an afterthought, is the same honesty principle as faithfulness applied one step earlier.
Does it work without Claude Code CLI?
Yes — on Claude Desktop with Skills enabled. Also works manually with any LLM by loading the markdown files into context (see "Use with other LLMs" below).
What's a research output look like?
See the example folder above. TL;DR: a folder with plan.md + sources/NN.md per source + findings/FN.md atomic theses + final <date>_<genre>.md report.
Every claim in the final report links to a specific sources/NN.md file.
Why so many files? Isn't this overkill?
For a 5-minute "what's the latest X" question — yes. That's why shallow mode exists (5-7 sources, no sub-agents, ~15 min). The full machinery is for medium (1 hour) and deep (3 hours) when you need to actually use the output for a decision.
The file-per-source structure is the key reuse mechanism. A single research often informs 3-5 future researches because you can cite individual sources/NN.md directly.
Is this just prompt engineering?
It's structured methodology + curated catalog + reusable templates + automation.
- The 11-phase workflow forces discipline
- 460+ stat sources catalog is curated knowledge
- 105 reusable blocks compose any report shape
scripts/validate_phases.pymachine-checks phase completeness, not just style- Weekly auto-validation keeps the catalog alive
- 25+ upstream awesome-lists give infinite discovery layer
Prompts are an implementation detail, not the value.
Can I use this commercially?
Yes — MIT licensed. Use it, modify it, integrate it into products. Attribution appreciated but not required.
The methodology is portable. ~70% of content is LLM-agnostic markdown templates.
| Component | Claude-specific | Universal |
|---|---|---|
SKILL.md frontmatter |
✓ | — |
Sub-agent Explore type |
✓ | — |
| 11-phase workflow | — | ✓ |
| 105 report blocks | — | ✓ |
| 29 search channels | — | ✓ |
| 460+ stat sources | — | ✓ |
To adapt:
- Load
SKILL.md+ relevantreferences/*.mdinto the LLM's context - Replace sub-agent parallelism with separate chat sessions per subtopic
- Manage source files (
sources/NN.md) externally — LLM produces content - PRs welcome for
codex/,gemini/,local/adapters
Deepdive — скилл для Claude Code, превращающий «загугли это» в дисциплинированный 11-фазный процесс.
- 11 фаз workflow: Reframing → Genre & block selection → Plan → Capability Discovery → Plan-review gate → Поиск → Claims-ledger + триангуляция → Evidence-фильтр → Синтез + multi-angle red team → Verify → Refresh targets
- 6 жанров отчёта: qa / explainer / decision / landscape / validation / custom
- 105 блоков в 10 категориях — переиспользуемые секции с шаблонами и анти-паттернами
- 29 каналов поиска с paywall fallback протоколом (включая api-direct)
- 460+ статистических источников в 14 cross-industry + 19 отраслевых категориях
- 39+ API endpoints для programmatic доступа (free no-auth приоритетны)
- plan.md с 17 секциями для прозрачности
- Plan-review gate (фаза 3.7) — единственная human-in-the-loop точка перед дорогой Фазой 4: план (гипотезы, жанр, каналы) показывается и утверждается ДО поиска; жёсткость по режиму (deep — ждать «Ок», medium — soft, shallow — skip)
- Multi-angle red team из враждебных ролей (Skeptic/Contrarian/Gap-hunter) с триажем severity (обязателен для medium/deep)
- Evidence-фильтр (фаза 5.5) — CRAG-классификатор keep/drop по паре (тезис, источник) перед синтезом: наивная подача всего найденного снижает качество (Search-o1 33%→24%), в синтез идут только relevant-цитаты из
evidence/ - Faithfulness-верификация (фаза 6.5, второй слой) — помимо liveness (жива ли ссылка) проверяется entailment «источник ⊨ тезис» (RAGAS-декомпозиция + ALCE), вердикты SUPPORTED/PARTIAL/UNSUPPORTED в
.verify/faithfulness.json - No-progress circuit breaker (фаза 4) — 2 раунда подряд без новой информации → стоп, нерешённое уходит в Open Questions вместо сжигания бюджета
- Дешёвый goal-check (фаза 4) — Haiku между раундами помечает каждый подвопрос met/partial/unmet, направляя следующий раунд и удешевляя дорогую Opus-оценку
- Least-to-most декомпозиция (фаза 4) — многошаговые подвопросы («X учитывая Y») раскладываются по уровням L1→L2 с накоплением контекста между ними вместо плоского параллельного запуска
- Phase-gate валидатор (
scripts/validate_phases.py) — машинная проверка, что каждая обязательная для режима фаза оставила артефакт; пропущенная фаза не проходит проверку - Weekly auto-validation через GitHub Actions
git clone https://github.com/Socialpranker/deepdive.git ~/.claude/skills/deepdiveТриггеры: «проведи ресёрч», «изучи тему», «копни глубоко», «deep dive»
Каталог растёт через PRs. Самые ценные — новые источники в stat_sources/ и api_sources/. См. CONTRIBUTING.md.
Built by Socialpranker · MIT License · Roadmap
If this skill saves you time, give it a star — it's the only metric I check.