Research and verify multi-source reports with citations

Go In Depth is a research harness that fans out web searches, fetches sources, adversarially verifies claims and synthesizes a cited report. It is run as a workflow named "go-in-depth" and is built for deep, multi-source, fact-checked research on any topic.

The workflow runs five phases:

1. Scope: decompose the question into 5 search angles. The angles are chosen to suit the question's domain, for example broad, academic or technical, recent news, contrarian or skeptical, and practitioner angles for general questions, or anatomy, common causes, serious differentials, authoritative references and red flags for a medical one.
2. Search: 5 parallel web search agents, one per angle, each returning the top 4 to 6 results ranked by relevance to the original question while skipping obvious SEO spam and content farms.
3. Fetch: deduplicate URLs (normalized by host without www and path without a trailing slash), fetch the top 15 sources, dropping any beyond that budget, assess each source's quality (primary, secondary, blog, forum or unreliable), and extract 2 to 5 falsifiable claims from each, every one supported by a direct quote and rated central, supporting or tangential.
4. Verify: a 3-vote adversarial verification per claim. Each voter is told to be skeptical and try to refute the claim, and it takes 2 of 3 refutations to kill it. Voters check whether the quote really supports the claim, look for contradicting evidence, weigh source quality against claim strength, check whether the claim is outdated, and watch for marketing claims or cherry-picked benchmarks. When in doubt, a voter defaults to refuting. Only the 25 highest-ranked claims are verified, ordered by importance and then source quality. A voter that skips or errors counts as an abstention, and a claim survives only when at least 2 valid votes came back and fewer than 2 of them refuted it, so an unadjudicated claim never slips into the report.
5. Synthesize: merge semantic duplicates, rank by confidence and cite sources. Confidence follows a simple rule: high for multiple primary sources with unanimous votes, medium for secondary sources or split votes, and low for a single source or blog-quality material. The report has an executive summary, findings with a confidence level (high, medium or low), the sources and evidence behind each, caveats, and a few open questions.

The result also lists refuted claims for transparency, with their vote counts, plus run statistics such as sources fetched, claims extracted, claims killed and URL duplicates. If every claim is refuted, the harness reports the research as inconclusive rather than inventing findings.

A basic invocation with a refined question looks like this:

Workflow({ name: "go-in-depth", args: { query: "best hybrid cars under $30k in the US for families" } })

Use it when you want a deep, multi-source, fact-checked research report on any topic, such as a product comparison (the skill's own example asks for the best hybrid cars under $30k in the US for families), a technical concept like how the transformer architecture handles positional encoding, or a claim to check, such as the efficacy of intermittent fasting for long-term weight loss in adults.

Before invoking it, check whether the question is specific enough. If it is underspecified, for instance "what car to buy" without a budget, use case or region, ask 2 to 3 clarifying questions to narrow the scope, then pass the refined question as the argument, weaving the answers in.

Invoke it through the workflow runner as Workflow({ name: "go-in-depth" }) for a bare run, or pass the refined question in args. Do not use it for quick facts. The stated limitations are slow execution (multi-agent searches, fetching and 3-vote verification take significant time), heavy context use (analyzing 15 full sources uses large context limits), and synthesis risk (it may struggle if the source material is weak or equally conflicting).

The input is a research question, passed as a string or as an object with a query field. If no question is provided, the workflow returns an error explaining how to pass one. The output is a structured report with the question, summary, findings, caveats, open questions, refuted claims, the list of sources with their quality, and run statistics.

The workflow relies on web search and web fetch tools inside a multi-agent workflow runner, and each phase returns structured output validated against a schema.

Anyone who needs a cited, cross-checked answer instead of a quick lookup: analysts, researchers and decision makers who can wait for a slower run in exchange for claims that survived adversarial verification.