Decision-grade global research with provenance.

The problem

Ask a general-purpose model for market research and it will happily produce a confident-sounding report (market size, growth rate, competitive landscape) built from a mix of real recollection, plausible extrapolation and, not infrequently, invented specifics that read exactly like the real ones. For a decision with money behind it, an unreadable mix of fact and fabrication is worse than no research at all, because it looks finished.

Why not just ask for it in the prompt

Asking a model to "research market X and cite your sources" does not reliably produce independent, verifiable sources — it produces citations that sound plausible, some of which do not resolve to a real, retrievable page. GlobalDive's source-grading and live-URL verification exist specifically because that failure mode is common enough to need a structural check, not a politer prompt.

How it works

• —Builds at least three competing hypotheses before searching, so research is not steered toward confirming a single assumed answer.
• —Runs geography-routed queries across a source registry — over 1,600 sources logged across 63 country sections in its own reference material (the skill’s own description claims 2,200+; that figure was not independently verifiable in the shipped package).
• —Runs a validation script against the finished report before export, and will not release a report with unresolved blocking errors.

Evidence and provenance

• —Grades every source A to D for credibility, and requires at least one A/B/C-level source behind every core conclusion.
• —Every citation must trace to a named, retrievable URL — a source that cannot be produced is not cited.
• —When public figures for the same metric conflict, GlobalDive shows the competing values rather than forcing a single number.

What it produces

• —A ten-section report: executive summary, industry definition, policy and macro environment, market size, demand, value chain, competitive landscape, business model, opportunities and risks, and a conclusion — plus a source-audit appendix.
• —A single thesis sentence naming the real opportunity, the entry mode, and the capabilities that would decide it.

What it must not do

• —If evidence is weak, states it plainly rather than presenting a guess as a finding.
• —Never calculates a market-share estimate as an arbitrary fraction of total market size.
• —Never includes raw search logs, chat-style prompts or local file paths inside a finished report.

In practice

A scan of a new market for a trade platform surfaces two very different figures for the same addressable market, from two sources that ultimately trace back to one original report re-cited by several outlets. Rather than averaging or picking the more convenient number, GlobalDive treats those as one source lineage, not independent confirmation, and shows both figures with their provenance rather than silently converging on one.

Why a reusable skill, not a one-off prompt

A one-off research prompt has no memory of the evidence-grading rules from the last one — the discipline has to be re-specified every time, and quietly degrades under time pressure. Packaging the source hierarchy, the hypothesis-first workflow and the validation gate as a skill means the same standard applies whether the research is urgent or not.

Mohammed's contribution

Defined the source-credibility hierarchy (A–D grading) and the pre-search hypothesis requirement, and set the validation gate that blocks a report from shipping with unresolved sourcing errors.

Notes

• —Ships in four functionally identical variants for Claude, Codex, Hermes and OpenClaw.