Deepdive

A structured meta-research skill for Claude Code

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) + decision spec: what you'll do with the answer
         ✓ 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, provenance-checked
         ✓ Ran adversarial pass (3 counter-arguments + an "execute this" role)
         ✓ Report ready: research/postgres-replication-vs-cdc/2026-05-21_decision.md
         ✓ Walked through your decision forks — you picked CDC tooling, logged to application.md

What this is

A Claude Code skill that turns "research this topic" into a 13-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.

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 plan.md documents every choice

Each source = file with verbatim quotes

Mandatory adversarial pass + opposition queries

Atomic theses in findings/FN.md reusable

Every claim → [s12] link → specific quote

Install in 30 seconds

git clone https://github.com/Socialpranker/deepdive.git \
  ~/.claude/skills/deepdive

That's it. Now type any of these in a Claude Code session:

"Investigate X"
"Изучи тему"
"Validate this hypothesis"

# 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 Skill

The 13-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.

Full instructions →

How it works

The skill runs 13 phases in order:

Phase	Name	What happens

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 / unmet with 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 → L2 instead 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 / UNSUPPORTED verdicts 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.

A well-cited report that changes nothing is still a failure — the skill's answer to that is a decision spine running through the whole pipeline, not a bolt-on question at the end:

Reframing (1) captures a decision spec, not just a topic: the action you'll take, who reads the report and what they do next, and at least one falsifiable if-then fork ("if the research shows X, I do A"). No fork means the research changes nothing — the skill downgrades to shallow "curiosity mode" with an explicit label instead of quietly running a full pipeline for a dead-end doc.
Sourcing (4-5) tracks provenance, not just source count. A root: field on every source flags what it's actually retelling — ten articles quoting one press release are one voice, not ten, so triangulation now requires ≥2 distinct roots in addition to ≥3 sources and ≥2 types. A snowball pass also chains citations backward (to the primary study everyone's retelling) and forward (who's citing it since), catching sources no keyword query would surface.
Synthesis (6) produces memo.md — a one-page decision memo (recommendation, resolved forks, 3 key numbers with sources, the main risk, next actions) designed to be the thing that actually enters your decision process, with the full report as its evidence base. A conditional recommendation mapped to your forks is required; an unconditional "it depends" is not an acceptable ending.
Red team (6) gets a fourth role — the Executor — who plays your own decision-spec consumer and tries to actually act on the report using only its content, flagging every place a hedge phrase or a missing number blocks a real decision.
Decision walkthrough (8) is mandatory at every depth, including shallow. The report isn't discussed, it's executed: the skill walks you through each fork one at a time ("research showed X [s03][s11] — does fork A fire? your call?") and logs the outcome — decided, blocked on missing data (one bounded gap-search, then honest blocked), or deferred — to application.md. A global ledger tracks whether past research actually led anywhere, and Phase 1 reads it back on your next research to flag a pattern of dead-end runs.

Want to compare models head-to-head? The eval harness scores any run on 6 axes.

What's inside

106 Report Blocks

10 categories: FRAME · EXPLAIN · COMPARE · MAP · VALIDATE · ANALYZE · CLOSE · PEOPLE · NUMBERS · CONTEXT

Each block has its own template, anti-patterns, and composition rules.

Block library →

29 Search Channels

Named strategies with query patterns + paywall fallbacks:

web-general · academic · preprint-servers · code-github · forum-discussion · news-current · industry-reports · regulatory-legal · competitive-signals · data-statistical-gov · product-analytics · crypto-analytics · api-direct · and more

Channels catalog →

460+ Stat Sources

14 cross-industry + 19 industry categories. Each entry: URL · Type · Access · Quality · Limitations · Combine-with · Fallback.

Categories: gov_macro · companies_public · crypto · health · education · climate_env · science · 19 industries

Sources catalog →

Plus a signal registry: 1072 endpoints verified live and filed by what question they answer — parliamentary APIs, advisory feeds, climate and macro endpoints. Loaded one file at a time, outside the base context budget.

Registry →

6 Report Genres

Genre	When
`qa`	Open meta-research
`explainer`	"How does X work"
`decision`	"X or Y"
`landscape`	"Who's in this space"
`validation`	"Is X true"
`custom`	Hybrid, assembled per question

Genres →

47+ API Endpoints

Free no-auth APIs prioritized:

Semantic Scholar · OpenAlex · CrossRef · arXiv · DefiLlama · CoinGecko · HN Algolia · World Bank · SEC EDGAR · ClinicalTrials.gov · PubMed · GDELT · NSF Awards · NIH RePORTER · Grants.gov · EPO Linked Open Data

Patents and grants are first-class categories: patents/ (USPTO ODP, EPO OPS, EPO LOD, WIPO) and grants/ (NSF, NIH RePORTER, CORDIS, Grants.gov). Every entry carries a Verified: date — access terms rot, and a stale "free, no key" sends a run into a paywall.

Auth via env vars only — skill never asks for keys inline.

API catalog →

Weekly Auto-Sync

GitHub Actions cron validates all endpoints + discovers upstream additions:

HEAD-check 47+ APIs weekly
Scan public-apis & awesome-public-datasets
Auto-PR for dead endpoints
Reports committed to reports/ branch

Workflow →

Model Routing

Per-phase model selection — quality where it multiplies, cheap where it parallelizes:

Reframing / plan / adversarial → Opus
Sub-agent fan-out (search) → Haiku (cheap × N)
Synthesis → Sonnet/high

~$2 instead of ~$8 on a deep run, and higher quality on critical phases. Override with with all on opus / with cheap mode.

Routing →

Eval Harness

Compare research quality across models. Same question, different configs, scored on 6 axes:

Deterministic (script): citation integrity, source diversity, cost
Semantic (LLM-judge): accuracy, coverage, adversarial honesty

Weighted sum with a citation floor — hallucinated sources can't win on depth. Verdict = quality per dollar.

Eval →

Citation Check

check_citations.py resolves every source URL — dead OPEN links flagged as likely hallucinations; transport flaps marked UNKNOWN, not penalized.

Ignores env proxies (trust_env=False). --strict for CI.

Verification runs four layers: liveness (does the source exist), faithfulness (does it actually entail the claim it's cited for), qualifier preservation (does the report still say what the ledger said), and construct provenance (do the frameworks, taxonomies and named "laws" the report uses exist outside it). Verdicts land in .verify/*.json, one producer per file.

The fourth layer exists because the first three all join on claim_id — and a fabricated name has none. That is the largest measured class of generation defect in research agents (FINDER/DEFT: strategic content fabrication, 18.95% of errors), and it passes a citation check with every URL alive.

Numbers get two independent passes: check_number_provenance.py on origin (who produced the figure; does one value circulate across supposedly independent roots) and check_number_arithmetic.py on computation — every derived figure in numbers.csv is recomputed from its own declared formula + inputs, and share groups must sum to 100. A percentage computed in prose is otherwise never re-checked by anything.

Check →

Phase-gate Validator

scripts/validate_phases.py reads a finished run's mode: frontmatter and checks that every phase mandatory for that depth left its file artifact — plan.md, state.md, claims.csv, numbers.csv, outline.md, evidence/, .verify/*.json, the dated report.

A skipped phase fails the check (--strict for CI) instead of the model just asserting "done" — it's a finish-up blocker, not advice. Machine insurance against the one failure mode a markdown methodology can't fix by discipline alone.

It also checks three things a missing-file test cannot: that state.md is a rebuilt round window (Known/Gaps/Next, size-capped) and not an appended second transcript; that every triangulated/contested claim is mapped to a section in outline.md — the synthesis gap, where retrieval succeeded and the finding never reached the report; and that no unsourced construct sits in the memo or TL;DR.

Its own inputs are machine-built too: scripts/build_sources_csv.py generates sources.csv deterministically from sources/NN.md frontmatter (--check for CI drift), and eval/validate_structure.py enforces the caveat: field as a strict enum (- / vendor / self-reported / disputed:sNN) instead of free text, so it stays greppable.

Validator →

Example folder

Sample output for a typical decision-genre research:

research/<topic-slug>/
├── plan.md                              # 17-section plan
├── state.md                             # Round window, rewritten each round
├── sources.csv                          # Index with C/R/B scoring
├── claims.csv                           # Claim ledger + triangulation status
├── numbers.csv                          # Every figure the report stands on
├── outline.md                           # section → block → claim_id map
├── 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
├── .verify/                             # One producer per file, many consumers
│   ├── authority.json                   # Phase 5.5 — who may assert this
│   ├── citations.json                   # Layer 1 — liveness
│   ├── faithfulness.json                # Layer 2 — entailment
│   ├── qualifiers.json                  # Layer 3 — scope drift
│   └── constructs.json                  # Layer 4 — named-construct provenance
├── memo.md                              # One-page decision memo (always)
├── application.md                       # Decision walkthrough verdict (always)
└── 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.

Contribute

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`

Full contributing guide →

FAQ

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.

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).

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.

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.

It's structured methodology + curated catalog + reusable templates + automation.

The 13-phase workflow forces discipline
scripts/validate_phases.py machine-checks phase completeness, not just style
A decision spec + mandatory walkthrough (phase 8) ties every research to an actual action, not just a document
Weekly auto-validation keeps the catalog alive
25+ upstream awesome-lists give infinite discovery layer

Prompts are an implementation detail, not the value.

The most common failure mode of open-ended research isn't bad sourcing, it's a report nobody knows what to do with. Three things guard against that specifically:

Reframing won't accept a topic without a decision spec — an action, a consumer, and at least one falsifiable if-then fork. No fork, no medium/deep run; it downgrades to a labeled shallow "curiosity" pass instead.
Synthesis is banned from ending on "it depends" without mapping the condition to one of your forks, and a fourth red-team role tries to actually execute the recommendation using only the report, flagging every hedge and missing number that blocks a real decision.
Phase 8 (Decision walkthrough) is mandatory at every depth — the skill walks you through your forks one at a time and logs whether each was decided, blocked on data, or deferred to application.md. A global ledger tracks unapplied research across sessions so the pattern doesn't repeat silently.

Yes — MIT licensed. Use it, modify it, integrate it into products. Attribution appreciated but not required.

Use with other LLMs (Codex, Gemini, etc.)

The methodology is portable. ~70% of content is LLM-agnostic markdown templates.

Component	Claude-specific	Universal
`SKILL.md` frontmatter	✓	—
Sub-agent `Explore` type	✓	—
13-phase workflow	—	✓
106 report blocks	—	✓
29 search channels	—	✓
460+ stat sources	—	✓

To adapt:

Load SKILL.md + relevant references/*.md into 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

Star History

Deepdive — скилл для Claude Code, превращающий «загугли это» в дисциплинированный 13-фазный процесс.

Что внутри

13 фаз workflow: Reframing → Genre & block selection → Plan → Capability Discovery → Plan-review gate → Поиск → Claims-ledger + триангуляция → Evidence-фильтр → Сверка с вики → Синтез + multi-angle red team → Verify → Refresh targets → Decision walkthrough
6 жанров отчёта: qa / explainer / decision / landscape / validation / custom
106 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
47+ 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
Decision Spec (фаза 1) — решение+срок / потребитель→шаг / ≥1 опровергаемая if-then вилка; без вилки ресёрч ничего не меняет — честный даунгрейд в shallow вместо мёртвого deep-отчёта
Provenance-триангуляция (фазы 4-5) — поле root: у каждого источника + snowball-пасс по цепочкам цитирований; триангуляция требует ≥2 разных первоисточников, не только ≥3 источников разного типа (десять пересказов одного пресс-релиза — один голос)
memo.md (фаза 6, всегда) — одностраничный decision-меморандум: рекомендация без «it depends», разрешённые вилки, ключевые числа со ссылками, next actions — то, что реально входит в твой процесс принятия решений
Роль «Исполнитель» в red team (фаза 6, medium/deep) — симулирует потребителя из Decision Spec и пытается исполнить решение по одному отчёту, ловя обтекаемый язык и пропущенные числа
Decision walkthrough (фаза 8, обязательна всегда) — отчёт не обсуждается, а исполняется: скилл проводит по вилкам решения одну за одной и записывает вердикт (принято / заблокировано данными / отложено) в application.md; глобальный ledger отслеживает, приводят ли твои ресёрчи к решениям вообще
Weekly auto-validation через GitHub Actions

Установка

git clone https://github.com/Socialpranker/deepdive.git ~/.claude/skills/deepdive

Триггеры: «проведи ресёрч», «изучи тему», «копни глубоко», «deep dive»

Вклад

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.

Deepdive

At a glance

README

Deepdive

A structured meta-research skill for Claude Code

What this is

Install in 30 seconds

How it works

What's inside

106 Report Blocks

29 Search Channels

460+ Stat Sources

6 Report Genres

47+ API Endpoints

Weekly Auto-Sync

Model Routing

Eval Harness

Citation Check

Phase-gate Validator

Example folder

Contribute

FAQ

Use with other LLMs (Codex, Gemini, etc.)

Star History

Что внутри

Установка

Вклад

Built by Socialpranker · MIT License · Roadmap

Comments (0)