disconfirm
Make AI try to break its own conclusion before it hands you one
$ claude
> /disconfirm ▌
Overview
A Claude Code skill that forces grounded research and a cheapest-falsifying test before the AI commits to any consequential claim. Instead of confidently telling you what you want to hear, Claude has to state what would prove its claim false, go get real evidence, design the cheapest test that could kill the claim, and report honestly when the data proves it wrong.
Why it exists
This skill was born from an experiment that disproved itself. I built an elaborate seven-session, verify-gated AI pipeline — panels of judge models voting, deterministic test floors, retry logic, audit trails. Then I made it measure itself against one sharp test. The machinery didn't win. Two moves did all the actual work: getting real evidence instead of reasoning from memory, and running one cheap test built to prove the idea wrong. So I threw out the rest and kept those two. The whole lesson was that the elaborate apparatus was the part you could skip.
What it does
States the claim in one line plus the single observation that would prove it false — no unfalsifiable claims dressed up as facts
Gathers real evidence: reads the actual code, runs the search, cites the source — deliberately looking for contradictions
Designs the cheapest test that could kill the claim, with the pass/fail line written down before running
Reports what the data said — especially when the data proved the AI wrong. A disproved assumption is the success case
Ships with an optional hook that auto-invokes the discipline whenever your prompt smells like a comparison or benchmark
A real example
The first thing I did with this skill was point it at itself. I asked Claude to confirm that I created it. The easy answer is “yes, you did.” Instead, it went and read the session transcripts, found the exact message where I commissioned the skill and the write operation that produced the file three minutes later, and came back with something more honest:
Confirm that I created this skill.
Yes — in conception and command. But the file itself was typed by Claude, on your order, three minutes after you commissioned it. (Verified from session transcripts, not from memory.)
Install
One command from the repo root. It drops the skill into ~/.claude/skills/ — nothing gets overwritten without you doing it yourself.
$ git clone https://github.com/edcadet10/disconfirm $ cd disconfirm $ ./install.sh
More from the toolkit
Want AI that proves its work?
I build tools like this because I use AI for real work every day — and I help businesses adopt it the same way.