A mirror, not a consultant.
Makes thinking harder, not easier.
AI tools that think for you make your thinking worse. This one refuses to.
Research shows a r = -0.68 negative correlation between AI usage and critical thinking (Gerlich, 2025). 83% of LLM users can't cite from texts they just produced. Passive delegation to AI reduces cognitive engagement.
Using an AI to improve your thinking is like using a calculator to improve your arithmetic. Unless the tool forces you to do the cognitive work, it replaces the very thing it claims to augment.
This skill produces questions, never answers. It creates desirable difficulties (Bjork) — conditions that feel inefficient but produce deeper understanding. If it feels uncomfortable, it's working.
The skill refuses to proceed until you've written your own reasoning
Your thesis The tensions you see A provisional conclusion Your open questions
Minimum 100–150 words. This is the only thing that separates cognitive augmentation from cognitive replacement.
Five phases, from your reasoning to your answers
Write your thesis, tensions, conclusion, and questions. The skill won't start without this.
Cloud (judgment) or Clock (verifiable)? Clock problems get redirected to adversarial-verify.
Classify each element: deductive (from principles), inductive (from examples), or abductive (best explanation).
Apply 3+ techniques from the arsenal of 9. Forced variety — no repeats.
3 questions, 1 contradiction, 1 load-bearing assumption, 1 likely bias. No answers. Just the mirror.
After you respond, the skill mirrors what changed since Phase 0. Not judgment — awareness.
Each run uses at least 3, never repeating
Separate vanity metrics from actionable metrics. "If this number changed, would you do anything differently?"
Make implicit assumptions explicit. Your reasoning rests on foundations you haven't examined.
"It's 12 months from now and this failed. What went wrong?" Forces prospective hindsight (Klein).
What happens at 10x? At zero? At negative? Test the reasoning under extreme conditions.
What happens in 6 months when context has changed? What assumptions become stale?
What if the user wants the exact opposite? How much of the reasoning survives?
Are you measuring what you did or what changed? Most OKRs measure output disguised as outcome.
Construct the strongest possible version of the opposing argument. If you can't, you don't understand the problem.
Can this claim be proven wrong? If not, it's not useful. "We're uniquely positioned" — unfalsifiable.
Patterns that look like thinking but aren't
Foundations of reasoning that haven't been examined
Numbers that feel good but don't change decisions
"The market is big" is not a competitive advantage
Citing successes without counting failures
"We're uniquely positioned" cannot be proven wrong
Shipping features is not the same as creating value
Built on cognitive science, not heuristics
Gerlich, 2025 — r = -0.68 correlation between AI usage and critical thinking. The empirical basis for Phase 0.
mdpi.com ↗ DifficultiesBjork & Bjork — Conditions that feel inefficient produce better long-term learning. The theoretical basis for "harder, not easier".
bjorklab.psych.ucla.edu ↗ CHI 2025CHI 2025 Workshop — 56 researchers on the understand/protect/augment framework for AI tools for thought.
arxiv.org ↗ Pre-MortemKlein, 1998 — Prospective hindsight increases risk prediction accuracy by 30%. The basis for cognitive forcing.
gary-klein.com ↗ OriginFullone, 2026 — The article that inspired this skill. Phase 0 blocking and cognitive interaction design.
fullo.net ↗git clone https://github.com/fullo/claude-adversarial-thinking.git
cp -r claude-adversarial-thinking/skills/adversarial-thinking ~/.claude/skills/claude marketplace add fullo-plugins https://github.com/fullo/claude-plugins-marketplace
claude plugin install adversarial-thinking@fullo-pluginsWorks with any tool that supports the Agent Skills format.
Start with /adversarial-thinking or ask naturally
/adversarial-thinkingchallenge my thinking on this strategypoke holes in my business planstress test my ideaare my OKRs measuring the right things?play devil's advocate on this decisionFor verifiable artifacts (code, data, schemas, docs, tests), use the companion skill
Adversarial Verify →