adversarial-thinking

Adversarial Thinking

A Claude Code skill that challenges your reasoning using Socratic method, cognitive forcing strategies, and desirable difficulties. It makes thinking harder, not easier.

Philosophy

This skill is NOT a consultant. It doesn't analyze your strategy, optimize your OKRs, or rewrite your pitch deck. It's a mirror — it reflects your reasoning back at you with uncomfortable questions.

The research is clear: passive delegation to AI reduces critical thinking (Gerlich 2025, r = -0.68). The only way an AI tool can improve thinking is by refusing to think for you and creating conditions where genuine cognitive effort is unavoidable.

Core rules:

• You must write your reasoning BEFORE the skill does anything (Phase 0)
• The skill produces questions, never answers
• If it feels uncomfortable, it's working

How it works

Phase 0 — Your thinking first (blocking): You must provide 100-150 words with your thesis, tensions, provisional conclusion, and open questions. The skill refuses to proceed without this. This is the only thing that separates cognitive augmentation from cognitive replacement.

Phase 1 — Classify the problem: Cloud (complex, judgment-dependent) or Clock (verifiable, decomposable). Clock problems get redirected to /adversarial-verify.

Phase 2 — Reasoning mode: Deductive, inductive, or abductive. Most strategic thinking is abductive — you observe patterns and construct narratives. The skill surfaces the alternative narratives you're not considering.

Phase 3 — Cognitive forcing techniques: 9 techniques, forced variety (min 3 per run, never repeat).

Phase 4 — The Mirror: 3 uncomfortable questions, the contradiction in your reasoning, the load-bearing assumption, the cognitive bias most likely at play. Ends with: "What are your answers to these questions?"

Phase 5 — Reflection Loop: After you respond, the skill mirrors changes between your Phase 0 and your current thinking. Not as judgment — as awareness.

Techniques

#	Technique	What it does
1	Noise vs Signal	Separate vanity metrics from actionable metrics
2	Assumption Excavation	Make implicit assumptions explicit
3	Pre-Mortem (Klein)	"It failed. What went wrong?"
4	Scale Shift	What at 10x? At zero?
5	Time Travel	What in 6 months? In 5 years?
6	Requirement Inversion	What if the user wants the opposite?
7	Outcome vs Output	Are you measuring what you did or what changed?
8	Steel Man	Strongest version of the opposing argument
9	Unfalsifiability Check	Can this claim be proven wrong? If not, it's not useful

Install

Copy the skill directory:

cp -r skills/adversarial-thinking ~/.claude/skills/

Or clone and copy:

git clone https://github.com/fullo/claude-adversarial-thinking.git
cp -r claude-adversarial-thinking/skills/adversarial-thinking ~/.claude/skills/

Or install via Claude marketplace:

claude marketplace add fullo-plugins https://github.com/fullo/claude-plugins-marketplace
claude plugin install adversarial-thinking@fullo-plugins

Usage

In Claude Code, type:

/adversarial-thinking

Or ask naturally:

"challenge my thinking on this strategy"
"poke holes in my business plan"
"stress test my idea"
"sfida il mio ragionamento"
"are my OKRs measuring the right things?"
"play devil's advocate on this decision"

What it challenges

• Buried assumptions — foundations of reasoning that haven't been examined
• Vanity metrics — numbers that feel good but don't change decisions
• Moats disguised as market size — "the market is big" is not a competitive advantage
• Unit economics at wrong scale — CAC that only works at a scale you haven't reached
• Survivorship bias — citing successes without counting failures
• Internal inconsistency — slide 3 says X, slide 7 assumes not-X
• Unfalsifiable claims — "we're uniquely positioned" cannot be proven wrong
• Output disguised as outcome — shipping features is not the same as creating value
• Overconfidence — certainty that exceeds the evidence

Companion skill

This skill handles cloud problems (complex, judgment-dependent, no single right answer). For clock problems (verifiable artifacts — code, data, schemas, documentation, tests), use adversarial-verify.

Compatibility