Deep Research

Use this skill to design and run repeatable research workflows that gather evidence across many sources, preserve provenance, and synthesize results into decision-ready outputs.

This skill covers both native deep-research agents (ChatGPT Deep Research, Gemini Deep Research, Perplexity Deep Research, Claude with web search) and custom agentic research pipelines (planner / searcher / verifier / synthesizer split).

Quick Reference

Question	Default
When to use a native agent vs custom pipeline?	Native for ad-hoc, open-ended questions. Custom for repeatable, auditable, or multi-source workflows.
What is the first artifact of any research task?	The source ledger — never the synthesis.
When is a source trustworthy?	When it is a primary document with a stable URL, author attribution, and a verifiable date.
What stops an unbounded research loop?	An explicit stop criterion defined before the loop starts (saturation condition or max iterations).
How to handle contradictory sources?	Separate them into evidence tiers; do not resolve by averaging.
How do I score a source?	Grade reliability and credibility separately — see references/evidence-packaging.md.
How do I evaluate or benchmark a research agent?	BrowseComp (arXiv 2504.12516: 1,266 hard-to-find-fact problems; Table 3 on p.6 reports an early OpenAI Deep Research at 51.5% accuracy vs 1.9% for GPT-4o with browsing, and footnote 1 notes Deep Research was trained on data that teaches BrowseComp-style tasks, a contamination caveat) and DeepResearch Bench (arXiv 2506.11763: 100 PhD-level tasks across 22 fields, scoring report quality and citation accuracy). Check the current leaderboards yourself; use ai-evals for judge reliability.

Use This Skill When

• You need to produce a sourced comparison, brief, memo, or research dossier.
• You need to choose between a native deep-research agent and a custom pipeline.
• You want to build a repeatable, auditable research workflow with provenance.
• You need to detect hostile sources, citation laundering, or model-output-as-source.
• You need to run a verifier subagent that has not seen the researcher's context.

Do Not Use This Skill For

• Single-lookup current facts — use ordinary browsing or direct tools.
• Long-lived retrieval serving — use ai-rag.
• Local note-vault packaging — use docs-notes-retrieval.
• Mining recent papers and blogs for methods to apply — use research-scout.
• Triaging new arXiv papers for a skill or product area — use research-arxiv-scout.
• Mining public GitHub repos for skills, team practices, or code patterns — use research-git.

Current Stance

The durable pattern is not "send a single query and summarize the top results."

The default architecture (as of last_validated date above):

1. Plan before searching: emit a structured research plan with queries, source targets, and success criteria before any tool call.
2. Ledger as contract: the source ledger is the primary artifact — synthesis is derived from it, not the other way around.
3. Verifier isolation: the verifier subagent must not share context with the researcher — it reads only the ledger and checks claims independently.
4. Evidence tier separation: primary sources (original docs, filings, specs) are never mixed with secondary commentary or model-generated summaries in the same evidence bucket.
5. Saturation-first synthesis: synthesize only after the research loop has reached saturation (new queries return no novel facts), not on a time budget.
6. Freshness windows are explicit: every claim in the final output carries a date anchor and a freshness class (stable / volatile / unknown).
7. Hostile source detection is mandatory for public-web research: adversarial SEO, AI-generated content farms, and citation-laundering chains are real threats.

Patterns

Full catalog → references/patterns-catalog.md

ID	Name	When to reach for it
P1	Source-ledger-as-contract	Always — every research task
P2	Plan-then-execute loop	Multi-query or multi-source tasks
P3	Multi-agent research swarm	Large corpora with parallel query paths
P4	Verifier subagent	Any claim that will reach a user or decision
P5	Freshness-window sourcing	Time-sensitive domains (AI, market, regulatory)
P6	Evidence-tier separation	When primary and secondary sources both exist
P7	Citation back-pointer	Every synthesis sentence that makes a factual claim
P8	Stop-criterion-as-eval	To prevent unbounded research loops
P9	Hostile-source detection	Public-web research, vendor comparisons
P10	Synthesis-after-saturation	All multi-pass research tasks

Anti-Patterns

Full catalog → references/anti-patterns-catalog.md

ID	Name	Why it matters
A1	Pre-synthesizing the answer	Frames all subsequent evidence gathering; produces confirmation bias
A2	Single-pass search	Misses contradictions, minority sources, and laterally relevant evidence
A3	Mixing evidence tiers	Primary docs and blog summaries in the same bucket degrade citation quality
A4	Model-output-as-source	LLM summaries carry hallucination risk; they are working notes, not source truth
A5	Unbounded research loop	Without a stop criterion, cost and time grow unboundedly with no quality signal
A6	Citation laundering	Citing a secondary article that in turn cites the primary — provenance chain breaks
A7	Verifier-with-same-context	Verifier that read the same sources as researcher adds no independence
A8	Vague consensus language	"Multiple sources say" without naming them is unfalsifiable
A9	Date-free volatile claims	Time-sensitive facts without date anchors decay silently
A10	Retrofitting citations	Writing first, citing later consistently inflates confidence in unsupported claims
A11	Overweighting recency	Newest is not most authoritative — prefer canonical primary sources
A12	Storing evidence with synthesis	Mixed buckets cause RAG follow-on to cite derived text as original evidence
A13	Treating the open-access corpus as the field	Paywalled prior work and unpublished negative results are silently absent; output skews toward hypotheses that only look promising

Decision Matrix: Native Agent vs Custom Pipeline

See full decision guide → references/native-deep-research-agents.md

Criterion	Native agent	Custom pipeline
Setup time	Minutes	Hours to days
Auditability	Limited (opaque search path)	Full (ledger + traces)
Repeatability	Low (non-deterministic)	High (seeded, logged)
Source control	None	Full
Freshness control	Model-dependent	Explicit per query
Multi-step planning	Implicit	Explicit (P2)
Verifier isolation	Not available	P4
Cost predictability	Low	High
Best for	Ad-hoc exploration, open questions, competitive overview	Repeatable workflows, regulated outputs, source-auditable deliverables

Decision rule: choose a native agent when speed matters and provenance does not. Choose a custom pipeline when the output will be cited, published, used in a product, or needs to be reproduced.

Native agents comparison

Full table with API surfaces and citation fidelity → references/native-deep-research-agents.md

Agent	API model ID / tool	Strengths	Limitations
ChatGPT Deep Research (OpenAI)	Legacy o3-deep-research / o4-mini-deep-research aliases — Responses API	Purpose-built source search and long synthesis	OpenAI's deprecations table lists the 2025-06-26 snapshots, with both aliases pointing at them, as shut down 2026-07-23 and names gpt-5.6-sol as the replacement. The separate model page can still describe o3-deep-research as available, so verify account access and rebuild the research loop against the replacement rather than assuming feature parity. Citations can go stale.
Gemini Deep Research (Google)	deep-research-preview-04-2026 / deep-research-max-preview-04-2026 (Interactions API)	Explicit plan review before execution, Google index, async + streaming, MCP support	Preview model IDs include date suffix — will change; in hands-on use (not benchmarked) synthesis depth ran thinner than ChatGPT DR
Perplexity Deep Research	sonar-deep-research (Chat Completions API)	Typically fastest in hands-on use (minutes, not tens of minutes), paragraph-level inline citations, real-time index	Citation laundering risk on SEO topics; shorter synthesis in hands-on use; model averages contradictions rather than flagging; multi-component pricing makes budgeting error-prone
Claude with web search	web_search_20260318 (current: adds response-inclusion control) / web_search_20260209 (dynamic filtering, GA) / web_search_20250305 (basic)	Full tool-call audit trail, structured output, P2/P4 composable, memory-tool-backed persistence	No dedicated DR mode; needs P2 loop for DR-equivalent depth
Grok DeepSearch (xAI)	web_search / x_search server-side tools (documented at docs.x.ai)	Only agent with X/Twitter as a primary indexed source; now has a documented, API-invocable tool surface	No dedicated report-generating deep-research model or endpoint comparable to the other four; citation auditability less mature

Architecture, Workflow, and Contracts

Full depth — planner / searcher / verifier / synthesizer subagent split, canonical contract shapes (ResearchPlan, SourceLedger, EvidenceTier, VerifierReport, SynthesisArtifact), stop-criterion design, freshness window classes — in references/agentic-research-loop-architecture.md.

Default Workflow

1. Frame the research question narrowly enough to be falsifiable.
2. Emit a structured research plan — see assets/templates/research-plan.template.md.
3. Execute the planner/searcher loop; record every source in the ledger before any synthesis.
4. Run the verifier subagent (must not share context with the searcher). The three scouts (research-scout, research-arxiv-scout, research-git) hand contested or high-stakes claims here from their Promotion Check; this verifier pass is their deep-verification step.
5. Separate evidence tiers: primary, secondary, and model-working-notes.
6. Check freshness windows; date-stamp all volatile claims.
7. Synthesize into the requested artifact; every factual claim carries a citation back-pointer.
8. Archive the ledger separately — see assets/templates/source-ledger.template.md.

Hypothesis Generation: The Co-Scientist Role Taxonomy

The planner / searcher / verifier / synthesizer split above answers a question. A different split is needed when the task is to generate and rank candidate ideas — research directions, product hypotheses, competing explanations for an observed effect — where there is no single retrievable answer to converge on.

Google Research's AI co-scientist (2025) decomposes that job into six specialized roles under a supervisor agent, each a separate context:

Role	Job	Why it is separate
Generation	Produce initial hypotheses from literature exploration plus simulated scientific debate between personas	Generation optimizes for coverage; mixing in evaluation collapses the candidate space early (see A1)
Reflection	Peer-review each hypothesis for correctness, novelty, and quality	A reviewer that also generated the idea cannot reject it — the same independence rule as P4
Ranking	Run an Elo-based tournament: hypotheses compete pairwise, ratings order the field	Pairwise comparison is more reliable than absolute scoring, and the rating is a live signal you can watch converge
Evolution	Refine the top-ranked hypotheses — simplify the concept, synthesize two ideas, or push into unconventional reasoning	Refinement is a distinct move from generation; it operates on winners, not on a blank page
Proximity	Compute a proximity graph clustering similar ideas	Surfaces where the field is crowded vs unexplored; prevents the tournament from ranking ten paraphrases of one idea
Meta-review	Synthesize patterns across all reviews and debates, feed them back as guidance	The only role that improves the system rather than a hypothesis

The loop is generate → debate → evolve, run iteratively: new hypotheses enter the tournament, top-ranked ones are evolved and re-enter, meta-review tightens the criteria each pass. Termination is the same discipline as P8 — stop on rating saturation (no new hypothesis breaks into the top tier for N rounds), not on a fixed round count.

How this composes with the rest of this skill. Generation and Reflection both need a ledger (P1) — a hypothesis grounded in nothing is a guess, and Reflection's novelty check requires knowing what already exists. Ranking is where evidence tiers (P6) do the work: a hypothesis supported by primary sources should outrank one supported by commentary, so feed the tier into the tournament rather than letting fluency decide. Meta-review is the natural place to catch A1 and A8 across a whole run.

Scope caution. This is a hypothesis-generation architecture. It orders candidate ideas by plausibility; it does not establish that any of them are true. The output of a tournament is a ranked queue for verification, never a finding — and see A13 for the literature bias that shapes what enters the queue at all.

Sources: Google Research, "AI co-scientist" (2025); described in Gulli, Agentic Design Patterns (Springer, 2025), Ch. 21 — Exploration and Discovery.

Research Stop Rule

Define the claim set and stopping rule before searching. Stop when every material claim has primary support or an explicit evidence gap, new searches yield no new source class or contradiction, and the contradiction ledger has an owner or bounded uncertainty statement. Continue when a load-bearing claim rests only on commentary, source independence is unclear, or a plausible counterclaim remains untested. Report saturation by claim and source class; elapsed time and citation count are not completion evidence.

Known Traps

Trap	Resolution
Starting with synthesis and backfilling sources	Freeze the ledger first; emit no synthesis until the ledger passes a minimum source count and verifier check
Using model outputs as sources in the next research pass	Stamp working notes with tier: model-output; the verifier rejects them as primary evidence
Verifier reads the same context as researcher	Give the verifier only the ledger JSONL, not the search session
Research loop never terminates	Define the stop criterion in the plan: saturation (no novel facts in N iterations) or hard cap (M total queries)
Citation laundering through secondary articles	Trace every citation to its primary document; reject chains longer than one hop without primary verification (same rule as A6 in the anti-patterns catalog)
Mixing old and new evidence for fast-moving topics	Apply P5 freshness-window sourcing; reject sources older than the window for volatile claims
Hostile SEO sources polluting vendor comparisons	Apply P9 hostile-source detection before adding any URL to the ledger
Session-state loss on agentic research re-runs	Persist the ledger and plan to disk before each iteration; resume from ledger state, not from conversation context
Trusting a vendor's feature guide as the source of truth for model availability	A provider's "how to use X" guide and its deprecations/sunset page are maintained on different cadences and can disagree — the guide may still recommend a model days before its documented shutdown. Check the deprecations page separately before wiring a pipeline to any specific model ID

• Check the provider's deprecations/sunset page, not just its feature guide, before committing a pipeline to a specific model ID — the two pages update on different cadences and the guide can lag behind an already-announced shutdown by days or weeks.

Navigation

References

• references/patterns-catalog.md — P1–P10 named patterns
• references/anti-patterns-catalog.md — A1–A13 with detection signals
• references/native-deep-research-agents.md — Comparison table, ChatGPT DR, Gemini DR, Perplexity DR, xAI Grok DeepSearch, Claude with web search, Anthropic memory tool, URL-health recipe
• references/agentic-research-loop-architecture.md — Planner / searcher / verifier / synthesizer subagent split
• references/research-workflow.md — Staged research loop and source-plan design
• references/evidence-packaging.md — Provenance, citation, and synthesis-packaging rules

Assets / Templates

• assets/templates/research-plan.template.md — Research plan scaffold
• assets/templates/source-ledger.template.md — Source ledger scaffold

Scripts

• python3 scripts/citation_verifier.py --input claims.jsonl — Checks JSONL of {claim, source_url, supporting_quote}. Fetches (or uses inline source_text for) each source, requires a normalized match of the quote against that text, detects negation, requires every number in the claim to appear in the quote with a compatible unit (a claim of 41.9 BLEU is not supported by a quote saying 28.4 BLEU), and downgrades to "needs_review" when a claim content word is absent from the quote ("children" vs "adults") or the quote adds a restricting qualifier the claim dropped ("only", "except", "limited to"); it reports "supported" only when all of that holds; a source that cannot be fetched is reported "unverified", never "supported". Every check is lexical (overlap, numbers, negation words, qualifier words), not meaning: a "supported" verdict on a claim that matters still needs a human or model entailment read. Exit 0 only when every claim is supported; exit 1 on any unsupported, unverified or needs_review claim (--allow-unverified and --allow-needs-review downgrade those two to reported-only); exit 2 on missing, empty or malformed input, including a line that is not a JSON object. Fetching allows only http/https to public addresses and re-checks every redirect. Tests: python3 -m unittest test_citation_verifier from scripts/.

Data

• data/sources.json — Curated source repos and tool references

Related Skills

• ../ai-rag/SKILL.md — Retrieval architecture and serving indexes
• ../docs-notes-retrieval/SKILL.md — Local note-vault packaging
• ../agents-mcp/SKILL.md — MCP search, docs, and memory tool surfaces
• ../qa-agent-testing/SKILL.md — Eval harnesses for productized research workflows
• ../ai-agents/SKILL.md — Agent architecture (planner/searcher/verifier split)
• ../research-scout/SKILL.md — Method mining from papers, blogs, and newsletters; hands contested claims here for verification
• ../research-arxiv-scout/SKILL.md — arXiv triage for a skill or product area; hands contested claims here
• ../research-git/SKILL.md — Repo mining for skills, practices, and code patterns; hands contested claims here

Learnings Loop

When prior decisions or pitfalls are relevant, consult learnings.consolidated.md if present; use learnings.md only for needed history or as the available fallback. Otherwise skip both.

After applying it, if you encountered a pattern worth remembering, a mistake worth preventing, or a domain fact that surprised you, append one dated bullet to learnings.md via agents-skills-feedback-loop/scripts/append_learning.py. Do not modify SKILL.md itself.