deep-research – skill Dữ liệu & Khoa học cho agent AI

Deep Research — Disciplined Meta-Research

Turn "research this topic" into an auditable, reusable investigation instead of a one-shot wall of text. The output is a folder you can return to in a month: every claim traces to a specific source file, the plan documents *why* each choice was made, and a refresh protocol lets you update it later without re-running everything.

This is the heavy, methodical end of research. It is not a fast overview — it is the workflow you reach for when getting the answer *wrong* costs more than the tokens spent getting it right.

How it differs from a quick research router

A router-style research skill (keyword-classify → delegate → short sequential search → markdown brief) is optimal when you need an answer fast and the decision risk is low. deep-research is the opposite trade: it pays for rigor. Use it when the answer feeds a strategy, an irreversible decision, a published artifact, or a hypothesis you need to actually test — situations where a shallow fallback would be a liability.

Concretely, deep-research adds what a fast overview does not: falsifiable hypotheses up front, parallel sub-agent fan-out across many channels, triangulation with explicit source-type diversity, a mandatory adversarial pass, per-source files with verbatim quotes, and a refresh_targets.md for delta-updates later.

The pipeline (9 phases)

Depth scales with the task — shallow runs the core phases inline; medium/deep add capability discovery, verification, and refresh targets.

| # | Phase | What it does |

|---|-------|--------------|

| 1 | Reframe | Rewrite the question, fix the underlying decision, state 2–4 *falsifiable* hypotheses |

| 2 | Genre & blocks | Pick the report genre (qa / explainer / decision / landscape / validation / custom) and its building blocks |

| 3 | Plan | Write plan.md: scope, structure, sourcing strategy, opposition queries, risk register, stop-criteria |

| 3.5 | Capability discovery | Audit available API keys/channels in the environment; map subtopics to sources; fall back to HTML where needed |

| 4 | Search (loop) | Dispatch sources → launch sub-agents in parallel → fetch & dedup → save each to sources/NN.md; re-evaluate between rounds |

| 5 | Score & triangulate | Rate every source on Credibility / Recency / Bias; require ≥3 independent, differently-typed sources per thesis |

| 6 | Synthesize + adversarial | Assemble the report from blocks, run 4 self-critique questions, add steel-manned counter-arguments |

| 6.5 | Verify | Lightweight citation check before closing |

| 7 | Refresh targets | Extract entities / numbers / hypotheses into refresh_targets.md — the entry point for future updates |

Core mechanisms

These are what separate a documented investigation from a confident guess:

• Triangulation. Every thesis must be backed by ≥3 independent sources of *different types* (primary / academic / industry / discussion). A claim with fewer is flagged "insufficient evidence," not stated as fact.
• Source-grounding. Each source becomes its own sources/NN_slug.md with metadata, verbatim quotes, and scores. No dangling claim — every assertion links back to a specific file. An empty fetch produces an empty claim, never a fabricated citation.
• Adversarial pass. Phase 6 always runs the strongest available reasoning: 4 self-critique questions plus an active search for counter-arguments and disconfirming evidence.
• Falsifiable hypotheses. Phase 1 commits to 2–4 hypotheses; Phases 5–6 explicitly confirm or refute each against the evidence, or mark it under-determined.
• Parallel sub-agents. Phase 4 launches search sub-agents concurrently (cheap models for broad web sweeps, stronger ones for reasoning-heavy subtopics) — never one-at-a-time.
• Refresh protocol. Phase 7 emits refresh_targets.md; an update <slug> run produces a delta (new entrants, entity changes, refreshed numbers, adversarial triggers) instead of replaying the whole investigation.
• Atomic findings. Reusable theses in findings/FN.md plus a sources.csv index — research compounds across questions instead of starting from zero each time.

Output structure

<root>/<slug>/
├── plan.md                  # scope, sourcing strategy, risk register, changelog
├── sources.csv              # index of every source with scores
├── sources/
│   ├── 01_<slug>.md         # one file = one source (metadata + verbatim quotes)
│   └── ...
├── findings/                # atomic, reusable theses (larger investigations)
│   └── F1_<short>.md
├── refresh_targets.md       # what to watch on update (medium/deep)
├── diffs/
│   └── YYYY-MM-DD_delta.md   # delta from an `update <slug>` run
└── YYYY-MM-DD_<genre>.md     # final report

When to use

• A low-quality answer is expensive: strategy, business plan, report, or article groundwork.
• Comparing N institutions, products, methodologies, or markets and you need defensible reasoning.
• Validating a hypothesis or a decision against external data.
• Meta-research: "understand how X works," "map the landscape of Y," answering a connected series of questions.

Anti-Patterns

• Don't skip the existing-work check. Before searching, see whether the answer is already in the project or in a prior research folder — you risk re-researching something you already have.
• Don't skip reframing, even when the request "seems clear." The decision behind the question usually changes the search.
• Don't output to chat only. Always persist sources and the report to files — the reuse value is in the folder, not the transcript.
• Don't fabricate citations. If a fetch returns nothing, the claim is empty — never invent a plausible URL. Bind every claim to a saved verbatim quote.
• Don't build conclusions on a thin corpus. Too few sources, or sources that all share one type, means triangulation hasn't happened — say so rather than overstating confidence.
• Don't skip the adversarial pass on medium/deep investigations. Confirmation-only research is the failure mode this skill exists to prevent.
• Don't run sub-agents sequentially. Fan-out in parallel; serial search wastes the wall-clock advantage.
• Don't collapse sources/ into one file. Per-source files are what make findings searchable and reusable across investigations.
• Don't pick the heaviest model for everything. Match model to subtask — cheap for broad sweeps, strong for synthesis and the adversarial pass.

Cross-References

• research router — for fast topic overviews where decision risk is low; deep-research is the heavyweight alternative when rigor matters more than speed.
• competitive-teardown — for comparing N competitors on a structured 12-dimension matrix.
• litreview / dossier / patent — domain specialists when the investigation is narrowly academic, person/company-focused, or patent-focused.