name: deep_research_skill description: Deep research harness — fan-out web searches, fetch sources, adversarially verify claims, then synthesize a cited report. Use when the user wants a deep, multi-source, fact-checked research report. If the question is underspecified, ask 2-3 clarifying questions to narrow scope first. version: 1.0.0 type: worker license: N/A category: research requirements: [cpu] metadata: {}

Deep Research Skill

A research procedure that searches multiple angles, extracts falsifiable claims, kills weak ones through adversarial self-verification, and synthesizes a cited report. Before invoking, check the question is specific enough; if underspecified (e.g. "what car to buy" without budget/use-case/region), ask 2–3 clarifying questions and weave the answers into the research question.

How to run (read first)

This is a single-agent procedure — you do every phase yourself, not by spawning other agents. Your engine is web_searcher_tool (already a deep, multi-engine search that fetches and synthesizes) plus web_fetcher_tool for reading a specific source. The five phases below are the method you follow solo: run several web_searcher_tool queries (one per angle), extract falsifiable claims, then critically re-check each claim yourself before keeping it. Finish with done_tool (result = the cited report).

The five phases — treat "Verify" and "Synthesize" as steps you do after the full claim pool is gathered.

Phase 1 — Scope

Decompose the research question into 5 search angles that approach it from different directions (don't just rephrase). Each angle becomes one search agent.

Phase 2 — Search (one web_searcher_tool query per angle)

For each angle, call web_searcher_tool. Aim for:

Search the web for the angle's query. Rank results by relevance to the

ORIGINAL question, not just the search query. Skip obvious SEO spam / content

farms. Include a short snippet capturing why each result is relevant.

Structured output only.

Then URL-dedup across angles and cap the number of sources fetched (≈top 15 by relevance), so duplicate and low-relevance URLs don't burn the fetch budget.

Phase 3 — Fetch + extract (per source)

For each novel source, read it with web_fetcher_tool and extract claims using this rubric:

## Source Extractor

Research question: "<QUESTION>"

Fetch and extract key claims from this source (URL, title, found-via angle).

1. Retrieve the page content.

2. Assess source quality: primary research/institution? secondary reporting?

blog/opinion? forum? unreliable?

3. Extract 2–5 FALSIFIABLE claims that bear on the question. Each must be a

concrete, checkable statement; include a direct supporting quote; and be

rated central/supporting/tangential.

4. Note publish date if available.

If the fetch fails or the page is irrelevant/paywalled, return claims: [] and

sourceQuality: "unreliable". Structured output only.

Collect all claims. Rank by importance (central > supporting > tangential), then by source quality (primary > secondary > blog > forum > unreliable), and keep the top N for verification.

Phase 4 — Verify (adversarial self-check)

For each ranked claim, deliberately try to REFUTE it yourself (use web_searcher_tool to look for contradicting evidence). Apply this checklist with a skeptical default — if it doesn't hold up, drop the claim:

## Adversarial Claim Verifier

Be SKEPTICAL. Try to REFUTE this claim.

Given the research question, the claim, its source + quality, and the

supporting quote, check:

1. Is the claim actually supported by the quote, or an overreach/misread?

2. Web-search for contradicting evidence — does any credible source dispute it?

3. Is the source quality sufficient for the claim's strength? (extraordinary

claims need primary sources)

4. Is the claim outdated? (old claims about fast-moving fields are suspect)

5. Is this marketing / press release / cherry-picked benchmark / forum

speculation?

refuted=true if: unsupported by quote / contradicted / low-quality source

for a strong claim / outdated / marketing fluff.

refuted=false ONLY if: well-supported, current, and source quality matches

claim strength. Default to refuted=true if uncertain. Evidence must be

specific. Structured output only.

A claim survives only if it was actually adjudicated: a quorum of valid votes AND fewer than the required refutations. Too many abstentions = unverified, which must NOT pass into the report. If every claim is refuted, report the research as inconclusive (sources may be low-quality or claims overstated).

Phase 5 — Synthesize

From the surviving claims (with their sources, quotes, and the evidence that held up under your verification), synthesize the report:

## Synthesis: research report

Given the confirmed claims (and, for transparency, the refuted ones):

1. Identify claims that say the same thing — merge them, combine their sources.

2. Group related claims into coherent findings, each directly addressing the

research question.

3. Assign confidence per finding: high (multiple primary sources, unanimous

votes), medium (secondary sources or split votes), low (single source or

blog-quality).

4. Write a 3–5 sentence executive summary answering the question.

5. Note caveats: what's uncertain, weak sources, time-sensitivity.

6. List 2–4 open questions that emerged but weren't answered.

Output

Return the report: the executive summary, grouped findings with per-finding confidence and citations, a transparency list of refuted claims with vote tallies, the source list with quality ratings, and run stats (angles, sources fetched, claims extracted/verified/confirmed/killed).