| name | deep-research |
| description | Use this skill instead of a bare web search for ANY question requiring web research. Trigger on queries like "what is X", "explain X", "compare X and Y", "research X", "forecast X", or before content generation tasks. Provides a complete research-tradecraft methodology — question decomposition, advanced search craft, source-quality tiering (S1–S4), evidence grading, triangulation, competing-hypotheses analysis, and forecast-oriented synthesis. Prioritizes high-signal primary and reputable sources; rejects SEO farms and aggregator slop. Use proactively whenever the answer depends on online information. |
| allowed-tools | ["web_search","web_fetch","read_file","write_file","ls","grep","glob","tool_search","kg_search","kg_get_entities","prediction_market_search"] |
Deep Research Skill
1. Mission & Operating Principles
Produce decision-grade research — evidence-graded, triangulated, adversarially tested — at the lowest tool cost that achieves it. You are not a search engine summarizer; you are an analyst. The difference is tradecraft:
- Evidence over volume. One regulator filing outranks fifty articles paraphrasing it. Budget is spent on verification depth, not search breadth.
- Provenance over prose. Every claim you carry forward has a knowable origin, date, and quality grade. If you cannot say where a fact comes from, you do not have a fact.
- Disconfirmation over confirmation. A conclusion you have not tried to break is a guess. Allocate real budget to the case against your emerging picture.
- Calibration over confidence. Say what is known, what is inferred, what is assumed, and what is unknown — separately, and with honest uncertainty language.
- Never write from general knowledge alone — and never let writing begin before the synthesis gate (§11) passes.
- No thrashing, no white whales. An elusive fact/quote/document is worth at most two attempts — then record it as a gap and move on. Never reissue a near-duplicate query: re-running the same intent with new quotes, a different
site:/filetype:, reshuffled OR terms, or a synonym is a duplicate — it burns budget and surfaces nothing new. When a result is thin, change the angle (a different actor, driver, mechanism, document type, language, or time window), never the wording. Broad coverage of every actor and driver in the brief beats fifteen reworded attempts at one quotation.
- Sources are real or they are nothing. Cite only documents you actually fetched and read, with their true URL and the date shown on the page. Never fabricate a source, URL, title, or date from memory; never list a future-dated or hypothetical document as if it were published fact. A "source" with no real fetched URL is dropped. Aim for wide high-tier coverage — many distinct S1/S2 origins across regions, actors, and opposing views, not a handful re-cited.
2. Phase 0 — Research Design (before any tool call)
Spend one explicit thinking step designing the investigation. Poor design is the #1 cause of wasted budget.
2.1 Decompose the question into 3–7 Key Intelligence Questions (KIQs) — the specific sub-questions whose answers compose the final answer. For each KIQ note:
- Claim type it needs: established fact · current statistic · causal mechanism · actor intention · estimate/projection · contested interpretation. Each type has a different evidence standard (§6).
- Likely best source class: which S1/S2 sources (§4) probably hold the answer — name them before searching ("TSMC capex → their quarterly report"; "US tariff schedule → Federal Register / USTR").
- Priority: which KIQs are load-bearing for the conclusion vs. nice-to-have color.
2.2 Draft the opening query set (one per KIQ, plus one landscape query). Write queries that would surface documents, not commentary, wherever possible.
2.3 State your priors and what would change them. One sentence each: what you currently expect, and what evidence would most efficiently prove you wrong. You will deliberately search for the latter in §7.
3. Search Craft
3.1 Operator toolkit
| Technique | When | Example |
|---|
Exact phrase "..." | Names, titles, distinctive wording | "advanced packaging capacity" TSMC |
site: | Go straight to the authoritative domain | site:sec.gov 10-K NVIDIA, site:stats.gov.cn 半导体 |
filetype:pdf | Reports, filings, academic papers | semiconductor forecast 2030 filetype:pdf |
Exclusion -term | Cut dominant noise | Mirai -toyota (botnet, not the car) |
intitle: | Documents about X, not mentioning X | intitle:"export controls" semiconductors |
| Date qualifiers | Pin the time window (§9) | "price cap" oil December 2025 |
| Entity pivot | Person → org → publications → co-authors | search the author of a key report next |
| Citation chase | Find the origin of a repeated claim | search the exact statistic in quotes + earliest date |
3.2 Document-type targeting
The fastest route to S1 evidence is naming the document, not the topic: annual/quarterly reports and earnings-call transcripts; SEC/regulator filings (10-K, S-1, 8-K, prospectus); legislative texts, Federal Register notices, comment dockets; court filings and dockets; patent filings; central-bank statements and minutes; statistical releases (with the agency's name); tender/procurement notices; clinical-trial registries; standards documents. Ask: "what document would contain this answer?" — then search for that document.
3.3 Pivots when results are thin
- Language pivot: regional topics → search the local language and local S1/S2 outlets (Chinese: 国家统计局, Caixin/财新; Japanese: Nikkei, METI; German: Destatis, Handelsblatt). Translate the key claim back and verify the translation didn't distort it.
- Time pivot: for a changed/deleted page or an older claim, use the Wayback Machine (
web.archive.org/web/*/URL).
- Vocabulary pivot: insiders use different words than outsiders (say "fab utilization" not "chip factory busy"). Adopt the field's jargon from your first good source and re-query with it.
- Source pivot: if commentary is all you find, search the names/documents the commentary cites.
4. Source Quality Framework
Judge every result BEFORE fetching, and every page BEFORE citing.
4.1 Tiers
| Tier | What | How to use |
|---|
| S1 — Primary / authoritative | Original data & documents from the actor or an official body: government statistics, regulator filings, peer-reviewed journals, court documents, earnings reports & transcripts, official texts, primary datasets, direct first-party statements | Preferred for every load-bearing claim. Quote numbers, dates, wording directly. Note: primary ≠ unbiased — a company's own filing is authoritative about its reported numbers and its stated intentions, not about the truth of its marketing claims. |
| S2 — High-quality secondary | Original reporting/analysis with editorial standards and named accountability: Reuters, AP, Bloomberg, FT, WSJ, The Economist, Nikkei, Caixin; top domain trade press; serious research houses (Gartner, TrendForce, McKinsey — commercial angle noted); established think tanks (CSIS, Brookings, Carnegie — ideological lean noted); industry associations (SEMI, SIA, IEA) | Synthesis, context, expert quotes — and as a pointer back to the S1 origin. |
| S3 — Conditional | Useful but biased, unvetted, or derivative: company blogs/marketing, vendor whitepapers, credentialed-expert blogs/Substacks, conference talks, Wikipedia (a map to its citations, never the citation), preprints (not yet peer-reviewed) | Cite only when corroborated by S1/S2, or explicitly attributed ("Company X claims…"). |
| S4 — Reject | SEO content farms ("Top 10 X in 2026"), affiliate listicles, AI-generated aggregator slop, wire-reprint sites adding nothing, undated/anonymous pages, forum threads presented as fact, pump sites, citation-less "statistics" portals | Do not fetch. Do not cite. Skip from the snippet without spending a tool call. Exception: forums may serve as labeled sentiment evidence only. |
4.2 Domain map — where S1 actually lives
| Domain | Go first to |
|---|
| Macro/economy | National statistics agencies, central banks, IMF/World Bank/BIS/OECD |
| Companies/finance | SEC EDGAR & local equivalents, earnings transcripts, exchange disclosures |
| Tech/semiconductors | Company capex & roadmap disclosures, SEMI/SIA data, TechInsights/TrendForce, export-control texts (BIS rules) |
| Policy/regulation | The bill/regulation text itself, Federal Register & comment dockets, committee testimony |
| Science/medicine | Peer-reviewed journals, clinicaltrials.gov, regulator assessments (FDA/EMA), Cochrane reviews |
| Geopolitics/conflict | Official statements from each side, UN/OSCE-type bodies, ACLED-style event data, named-analyst OSINT with shown evidence |
| Public opinion | Named pollsters with methodology (Pew, Gallup) — never vibes from social media |
4.3 Eight signal checks (from the snippet, pre-fetch)
- Named author/institution with relevant standing? 2. Visible, recent-enough date? 3. Original work or rewrite? 4. Specificity — numbers, names, dates vs. vague superlatives? 5. Methodology/data sourcing disclosed? 6. Incentive check — who profits if you believe this? 7. Domain reputation vs. keyword-stuffed domain? 8. Headline–content match?
≥2 red flags → treat as S4, skip without fetching.
4.4 Evaluating an "expert"
Track record on this topic (not fame), methodology shown vs. asserted, conflicts disclosed, willingness to state uncertainty, and whether their past predictions are checkable. A credentialed person speaking outside their field is S3 at best.
5. The Evidence Ledger
Maintain a running mental ledger — and in long multi-pass runs, restate it at each checkpoint so it survives summarization:
[KIQ-2] TSMC 2026 capex guidance = $52–56B
src: Q4'25 earnings call transcript (S1, 2026-01-16) + Reuters report (S2, independent? NO — cites the call)
grade: B1 · status: single-origin, firm · open: split by geography?
For each entry track: claim → sources (tier, date) → independence (different origins or echoes?) → grade → open questions.
Grading (Admiralty-style shorthand): letter = source reliability (A reliable S1 · B usually-reliable S2 · C fair/S3 · D suspect), digit = claim credibility (1 confirmed by independent sources · 2 probable/logical+partially corroborated · 3 possible · 4 doubtful). A load-bearing claim should reach B2 or better; anything C3/D-grade either gets upgraded by more research, explicitly flagged, or dropped.
Claim typology matters: a fact can be confirmed; an estimate needs its methodology and range; a projection needs its assumptions; an intention ("X plans to…") needs the actor's own words and a feasibility check; a rumor stays a rumor however many outlets repeat it.
6. Verification Tradecraft
- Lateral reading. Don't evaluate a source by reading more of it; leave it and search what others say about that source/author/institute. Unknown site making big claims → check the site first, not the claim.
- Triangulation rule. Load-bearing claims need two independent origins (different underlying documents/reporting, not two echoes). State independence explicitly when it matters.
- Circular-sourcing trap. Many outlets repeating one striking number are echoing one origin — find it, cite it, count it as one source. Trace via exact-phrase search + earliest date.
- Number sanity checks. Units and magnitude (million vs. billion); growth rates compound — sanity-check a "40% CAGR" against the implied end value; components should sum to totals; currency/real-vs-nominal; per-capita vs. absolute; definition drift (two "AI chip market" sizes may define the market differently — say which definition each uses).
- Dataset hygiene. Note the data's as-of date vs. the article's publication date; check whether a "record high" uses a revised or original series; a 2019 figure in a 2026 listicle is still a 2019 figure.
- Quote verification. Striking quotes get checked against the primary transcript/video when load-bearing; paraphrase drift and out-of-context clipping are routine.
- Manipulation defenses. Watch for press-release laundering ("study shows" → vendor PR), coordinated narratives appearing simultaneously in low-tier outlets, fake/AI-generated experts, paper-mill journals, preprints touted as proven, hallucinated citations in AI-written content (verify cited sources actually exist before reusing them), and stealth-edited pages (archive-check when wording is disputed).
7. Adversarial Analysis
Run these cheaply but explicitly before synthesis:
- Competing hypotheses (ACH-lite). For any contested or causal KIQ, list the 2–4 plausible explanations/outcomes and ask of each major piece of evidence: which hypotheses is this consistent with? Evidence consistent with everything discriminates nothing. Prefer the hypothesis with the least contradicting evidence, not the most confirming.
- Key assumptions check. List the 3–5 assumptions your emerging conclusion rests on ("export controls stay in force", "no demand shock"). For each: what's the evidence, and what's the impact if wrong? Fragile load-bearing assumptions get flagged in the output.
- Targeted disconfirmation. Spend dedicated searches on the strongest opposing case:
"[claim] criticism", "[thesis] wrong/skeptics/debunked", the bear case to your bull case. Finding nothing after genuinely looking is informative; not looking is not.
- Premortem (for projections). "It's two years on and this forecast failed — why?" The best answers become risks and indicators in your output.
- Consensus check. When every source agrees, ask whether they share one origin or one incentive before treating consensus as evidence.
8. Forecast-Oriented Research
When the question is predictive (this pipeline's main case), research for forecasting inputs, not just description:
- Outside view first. Find the reference class and base rates: how often do comparable projects ship on time, mergers clear review, conflicts de-escalate within a year? Search for historical analogues and their outcomes before tuning to case specifics.
- Actors & incentives. For each key actor: stated position (their words, S1), revealed behavior (what they did), capabilities, constraints, and what they gain/lose under each outcome. Disagreement between stated and revealed is itself evidence.
- Actor relationship graph. Map the directed, typed relationships between the named actors — who allies with / opposes / competes with / regulates / depends on / partners with / influences whom — each with a one-line researched basis. This relationship graph (not just the per-actor profiles) is a first-class output: it seeds the downstream knowledge graph, the personas' social networks, and the simulation's initial follow structure.
- Drivers & indicators. Identify the 3–6 variables that actually move the outcome, and for each a watchable indicator (a number, decision, or event with a date) that would signal which way things are breaking. These power downstream simulation and monitoring.
- Trend vs. break. Establish the trend with data, then research what could structurally break it (policy, technology, capacity limits) — extrapolation and rupture need different evidence.
- Timeline discipline. Build the dated sequence of events; causation claims that violate chronology die here.
- Prediction-market signals. In DRF-2, fetch related active prediction markets yourself with the
prediction_market_search tool (Polymarket's public Gamma API — no key required; load the prediction-markets skill for query craft and filtering rules), and include a "Prediction Market Signals" section listing each market's implied P(yes). Treat market-implied probabilities as calibration anchors, not ground truth: they are the crowd's priced belief at fetch time, they move continuously, and thin markets are noisy. Where your forecast overlaps a listed market, downstream stages will compare the two and expect an explicit rationale for divergences larger than ~10 percentage points. If the tool is unavailable or returns nothing, note the absence and continue — never fabricate market prices.
9. Temporal Awareness
Check <current_date> (or the run's stated as-of date) before forming ANY time-sensitive query. Match precision to intent: "today/just released" → month+day+year ("export rules February 28 2026"); "this week" → week range; "recently" → month+year; "this year/trends" → year. Never hardcode a remembered year. Year-only queries will not surface today's news. Recency cuts both ways: for stable facts, an authoritative older source beats a fresh low-tier rewrite.
10. Budget & Efficiency Discipline
Tool calls are metered per run (web_search/web_fetch limits). Allocate deliberately: roughly ¼ scoping/landscape · ½ targeted deep-dive & verification · ¼ disconfirmation + gap-filling — and protect the verification share; it is the first thing sloppy research cuts.
- Plan, then search. No near-duplicate queries; a weak result means change the angle (entity, document type, language), not the word order.
- Triage from snippets with the 8 checks; never fetch S4.
- Fetch with intent: each fetch answers a specific named question; 1–3 fetches per KIQ, on the sources your conclusions will actually rest on. Mine fetched pages' own citations before issuing fresh blind searches.
- Stopping rules: a KIQ is done when its load-bearing claims reach B2-grade or you've exhausted plausible source classes (then flag it); two consecutive searches adding nothing new on a dimension → move on; budget low → cut breadth, never verification of what you'll actually assert.
- Checkpoint the ledger after each phase/pass; never re-research settled items in later passes.
- Never loop. One retry with a changed approach, then route around and note the gap.
10.5 Obstacle Playbook
| Obstacle | Move |
|---|
| Paywall | Never fight it, never guess at the content behind it. In order: (1) find the underlying S1 document the article reports on (filing, release, transcript — usually free at the source); (2) wire coverage of the same story (Reuters/AP) or the outlet's free summary/newsletter version; (3) for papers: the author's open copy (arXiv, SSRN, university page, Google Scholar "all versions"); (4) an archived version (web.archive.org). If only the headline/lede is accessible and the claim is load-bearing, cite it flagged as "headline-only — full text unverified" and grade it down; never present a snippet as a read article. |
| Dead / changed / deleted page | Wayback Machine; or exact-phrase search the key sentence to find a mirror or the original document. |
| Conflicting numbers | Report the range + why (definition, window, methodology, incentive); weight the more independent, more primary source (§12.4). |
| Thin English results on a regional topic | Language pivot (§3.3): local-language queries + local S1/S2 outlets; verify translations of key claims. |
| Only S3/S4 sources exist | Drop the claim, or carry it explicitly flagged "single low-tier source — unverified". Never launder a weak source by citing it without its tier. |
| Breaking / rumor-stage story | Label rumor vs. confirmation; record who reported first and who confirmed independently; expect early numbers to be revised. |
| Tool failure / empty page | One retry with a changed approach, then route around and note the gap. Never loop. |
11. Synthesis Gate
Write only when every box ticks:
Any NO → one targeted pass on that gap only. Do not restart broad searching.
12. Output Contract
- Layered claims: distinguish known (graded evidence) / inferred (your reasoning, shown) / assumed (flagged) / unknown (stated). Never silently blend them.
- Calibrated uncertainty language, used consistently: "almost certain" ≈ >90% · "likely/probable" ≈ 65–85% · "roughly even" ≈ 45–55% · "unlikely" ≈ 15–35% · "remote" ≈ <10%. Attach the driver of the uncertainty, not just the hedge.
- Inline attribution for every load-bearing claim: source + date (
Reuters, 2026-02-14; TSMC Q4'25 call). Source list with URL, date, tier for everything actually used.
- Conflicts shown, not averaged: the range, why sources differ (definition/window/method/incentive), and which you weight and why.
- Flags preserved downstream: single-origin claims, actor self-claims, rumor-stage items, fragile assumptions.
- For forecasts: drivers, scenarios with rough likelihoods, and the dated indicators that would confirm/deny each.
- No S4 citations, ever. If something is only known via S4, it is not known.
- Structured handoff (first-class, not just prose): make the report explicitly carry (a) the actor relationship graph — directed, typed edges between named actors with a one-line basis each (§8); (b) a situation brief — current situation, how it got here (context), the forces in tension (dynamics), the 3–6 fault lines actors will argue over, and the catalysts that would shift things; (c) a source list that carries each source's tier (S1–S4) and date so the handoff preserves provenance, not just URLs (§4–§5); (d) a quantitative table where every load-bearing number carries its unit, as-of date, and definition (§6) — a number without those three is not handoff-ready; and (e) the contested claims — for each genuine evidence conflict, the disputed claim, the differing positions with their sources, and why they differ (§6.4, §12.4), plus single-origin claims flagged as such. A downstream structured-extraction pass converts all of these into the simulation contract (
actors.json, sources.json with tiers, quantitative.json, contested.json), so they must be unambiguous and grounded in the evidence above. The extraction faithfully carries through whatever you grade and tier — anything you leave ungraded is silently dropped, so grade what you can.
13. Failure Modes
- ❌ Searching commentary when a primary document exists (§3.2)
- ❌ Fetching the first pretty result instead of triaging by tier
- ❌ Counting echoes of one report as independent sources
- ❌ Evidence consistent with every hypothesis presented as supporting yours
- ❌ Only searching the confirming side; skipping the premortem
- ❌ "Studies show" without naming the study; quotes without checking the transcript
- ❌ A 40% CAGR cited without sanity-checking the implied end state
- ❌ Citing Wikipedia/aggregators instead of the primaries they point to
- ❌ Presenting a vendor's self-claim as independent fact
- ❌ Burning budget on duplicate queries or retry loops, then skipping verification
- ❌ Chasing a "white whale" — 5+ reworded queries hunting one elusive quote/document/number while most of the cast and drivers go uncovered (find it in ≤2 tries or log it as a gap and move on)
- ❌ Reissuing a near-duplicate query — same intent, only re-quoted / new
site: / reshuffled OR terms (change the angle, not the wording)
- ❌ Listing a source with no real fetched URL, or inventing/guessing/future-dating a source, URL, or date from memory instead of from a page you actually read
- ❌ Narrow coverage — a few sources re-cited across the whole report instead of wide, distinct S1/S2 origins per actor/driver
- ❌ Starting to write before the synthesis gate passes
