Adversarial Researcher: Try to Break the Claim

What it does

Researches a technical question by trying to prove it wrong. It states the claim so it can be false, names what evidence would kill it before searching, covers papers, patents, shipped products, and standards, and gives a verdict: supported, weakened, or killed.

Why it exists

Most research searches to confirm. It stops at the first reassuring result and misses the paper, patent, or product that already does the thing. This skill spends part of its budget looking for the counter-evidence, using the vocabulary a competitor or examiner would use.

How it works

1. Writes down the falsifiable claim and what would kill it before any search.
2. Picks a search budget by stakes: Quick (5 searches), Standard (15), or Deep (40), and stops early when the answer is clear. This keeps web-search costs predictable.
3. Uses a bundled source guide: where to search each layer (academic, patents, products, standards), patent query syntax and CPC classes, and how to weigh sources.
4. Runs a second pass aimed at breaking the conclusion, searching by mechanism rather than your own terms.
5. Double-sources the claims the conclusion rests on, and marks which sources were read in full.

What you get

The claim and its falsification test, a table of the closest existing work with dates and gaps, the strongest counter-evidence, a verdict with confidence, open questions, and dated sources.

When not to use it

Quick single-fact lookups or general news summaries.

Open source

Part of claude-rigor-skills: 30 MIT-licensed skills that make AI show its work. The demo below, with its inputs, is in the examples folder.

See it in action

Request:

Use the researcher skill: does anything already exist that does room-level indoor localization using only a phone's own speaker and microphone (ultrasonic chirps), with no installed beacons? Standard depth.

Result: the complete, unedited output from a fresh Claude Code session is in examples/researcher.

Known limitations

• Depends on the agent's web search tool. Paywalled papers and some patent databases can't be read in full; the report marks what was read directly and what came from summaries.
• Patent coverage through general web search is shallower than a dedicated patent database search; it says so when it applies.
• Results reflect what's publicly indexed at the time of the run.

FAQ

Why does it say my idea is "killed"? Because it found evidence that the claim, as stated, is false. It also lists the gaps the existing work leaves open, which is often where a narrower version of the idea still stands.

How much does a run cost? It's bounded by the search budget you pick. The Standard demo used 11 searches and 3 page reads.

Can I trust the sources? It weighs sources (peer-reviewed and primary over marketing pages), double-sources the claims the conclusion depends on, and marks anything single-sourced.

Compatibility

Works with any agent that supports the SKILL.md standard and has web search. The demo was run with Claude Code 2.1.296 (claude-sonnet-5-5) using its WebSearch and WebFetch tools.

Documentation: https://github.com/hemapriyan-rk/claude-rigor-skills/tree/main/skills/researcher