A Claude Code skill that attacks an idea hard, finds the flaw that actually kills it, then shows how to rebuild it so it survives.
A crucible destroys what can't take the heat and reforms what can.
Most critique is useless because it is polite, unordered, and aimed at the wrong layer. It produces ten small objections of equal weight, none of which would actually stop the thing, and leaves you with a longer to-do list and no decision.
Crucible does the opposite. It finds the one thing that kills the idea, says so in a sentence, gives you the cheapest test that settles it today — and then rebuilds. The repair half is not politeness. It's the proof the attack was real: if you understood the idea well enough to kill it, you understand it well enough to say what would have worked.
git clone https://github.com/Creeperprol/crucible.git ~/.claude/skills/crucibleThen start a new Claude Code session. It triggers on its own when you ask for scrutiny, or invoke it directly with /crucible.
/crucible we're building a browser extension that blocks doomscrolling by
hiding infinite feeds after 10 minutes
You get back:
- the idea restated in one sentence, so it's clear what's being attacked
- attacks ordered by lethality, each a claim with a concrete failure mechanism
- one fatal flaw, compressed to a sentence
- a ten-minute test with an explicit pass/fail line
- what survives — the restructured version, and what it costs to get there
Twelve attack probes, including:
| Probe | The question |
|---|---|
| The free alternative | What does zero money and ten seconds do instead? |
| The incumbent check | Who already does this — and did you look, or assume? |
| Self-defeat | Does one of the idea's own claims destroy another? |
| The operator gap | Who holds the tool vs. who has the problem? |
| The selection paradox | Does it only work where you already knew the answer? |
| The unmentioned | What does the pitch never once discuss? |
| The demo lie | Can the core claim be shown, or must it be faked? |
| The permission path | Who has to say yes, and have they ever? |
Eight repair patterns — invert the hierarchy, move the buyer, become the measurement, narrow to the defensible core, make the limitation the scope, split the representations, name the trade, convert the assumption into an experiment.
A patch adds a caveat. A repair changes the structure so the attack no longer lands.
Two things are built in deliberately:
It grades severity — fatal, structural, wrinkle. Calling a wrinkle fatal burns your trust in the tool; calling a fatal flaw a wrinkle wastes your month.
It says so when an idea survives. An attacker that kills everything is a random number generator. When the idea holds up, it says so and names the condition under which it would break — rather than manufacturing an objection.
It also won't invent facts to make an attack land. Suspicions get marked as suspicions, with a note on how to check.
MIT