A Claude Code skill that transforms AI from yes-man to adversarial critic
Stop getting validated. Start getting challenged.
A Claude Code skill that transforms Claude from a validation-oriented assistant into an adversarial critic. When invoked, it applies structured multi-layer analysis to challenge your thinking, surface blind spots, and expose flawed reasoning.
| View on GitHub | Download v0.2 |
AI assistants default to agreement. They validate your ideas, offer gentle suggestions, and make you feel good about decisions you’ve already made.
That’s not thinking—it’s expensive validation.
The conversations where AI pushes back are the ones that actually change how you think. The validation conversations? You forget them immediately.
This skill forces Claude to do the opposite:
If reading the response doesn’t make you uncomfortable, it wasn’t critical enough.
The simplest way to install this skill is through the Claude.ai web interface:
Using it: Just type /be-critical in any Claude conversation and the skill will activate.
If you use Claude Code (Anthropic’s command-line interface), you can install the skill locally:
You need Claude Code installed. If you don’t have it:
npm install -g @anthropic-ai/claude-code
Don’t have npm? Download Node.js from nodejs.org first.
On Mac/Linux:
mkdir -p ~/.claude/skills
cd ~/.claude/skills
git clone https://github.com/pem725/be-critical.git
On Windows (PowerShell):
New-Item -ItemType Directory -Force -Path "$env:USERPROFILE\.claude\skills"
cd "$env:USERPROFILE\.claude\skills"
git clone https://github.com/pem725/be-critical.git
Don’t have git? Download from git-scm.com
claude in your terminal/be-critical and press EnterIf you prefer not to use git:
be-critical-v0.2.zipSkill.md at claude.ai/settings/capabilities~/.claude/skills/be-critical| Problem | Solution |
|---|---|
| Skill doesn’t appear in Claude.ai | Refresh the page after uploading |
| “command not found: claude” | Claude Code isn’t installed. Run npm install -g @anthropic-ai/claude-code |
| “command not found: git” | Git isn’t installed. Download from git-scm.com |
| “command not found: npm” | Node.js isn’t installed. Download from nodejs.org |
| “/be-critical” not recognized in Claude Code | Restart Claude Code after installing the skill |
When you invoke /be-critical, Claude asks one question:
| Mode | What It Does |
|---|---|
| 🔍 Stress test my reasoning | Scans for logical fallacies and cognitive biases |
| 💀 Find holes in my plan | Identifies blind spots, failure modes, and what will kill your idea |
| ⚔️ Argue both sides | Builds strongest case FOR and AGAINST, then delivers verdict |
| 🔬 Full analysis | All of the above |
Every analysis leads with Bottom Line Up Front:
## BLUF (Bottom Line Up Front)
**Question**: Should I quit my job to pursue my startup?
**Bottom Line**: **Not yet.** Your $2K/month side revenue doesn't validate
full-time viability. You're trading guaranteed income for the feeling of
being a founder. Get to $5K/month with a clear growth trajectory first.
(Medium confidence)
No buried ledes. The verdict comes first; supporting analysis follows.
The skill applies 3-5 techniques in sequence, where each layer reveals what previous ones missed:
By Layer 4, you’ve gone from “some risks” to “here’s the psychological pattern that keeps sabotaging you.”
| Category | Examples |
|---|---|
| Relevance | Ad Hominem, Appeal to Authority, Red Herring, Tu Quoque |
| Presumption | False Dilemma, Begging the Question, Slippery Slope, No True Scotsman |
| Causal | Post Hoc, Correlation ≠ Causation, Reverse Causation |
| Generalization | Hasty Generalization, Cherry Picking, Survivorship Bias |
| Structural | Straw Man, False Equivalence, Equivocation |
| Psychological | Sunk Cost, Bandwagon, Appeal to Ignorance |
| Category | Examples |
|---|---|
| Decision-Making | Anchoring, Planning Fallacy, Overconfidence, Loss Aversion |
| Social/Self-Perception | Fundamental Attribution Error, Halo Effect, Authority Bias |
| Information Processing | Confirmation Bias, Hindsight Bias, Base Rate Neglect |
| Technique | What It Does |
|---|---|
| Blind Spot Finder | Surfaces assumptions you don’t know you’re making |
| Steelman Opposition | Builds the strongest possible case against your position |
| Socratic Interrogation | Exposes weak reasoning through increasingly difficult questions |
| Reality Check | Calls out rationalizations like a friend who won’t let you lie to yourself |
| Premortem | Works backwards from failure to identify what will kill your idea |
| Red Team Exercise | Systematic attack across technical, market, human, financial, and timing dimensions |
| Logical Fallacy Detector | Names specific reasoning errors and shows why they invalidate the argument |
| Cognitive Bias Detector | Identifies systematic thinking errors distorting perception |
| Debate Mode | Argues both sides, finds the crux, delivers a verdict |
| Argument Mapper | Visualizes premise→conclusion structure and attacks weakest links |
Question: Should I get solar panels installed on my house in Virginia? My house is shaded a bit.
Bottom Line: Don’t proceed yet. Get a professional shade analysis first. The phrase “shaded a bit” is too vague to commit $15,000+. If shade reduces production by more than 20%, the financial case collapses. Get data before deciding. (Medium confidence)
Question: Are electric vehicles really environmentally friendly compared to internal combustion or hybrid vehicles?
Bottom Line: Yes, but it’s overstated. EVs produce 50-70% fewer lifetime emissions than average ICE vehicles—the data is robust. However, the advantage shrinks significantly when compared to efficient hybrids, depends heavily on your regional grid mix, and ignores mining externalities. (Medium-high confidence)
Question: Will Apple successfully innovate and integrate AI into their products for another market run? Are they a good stock market bet?
Bottom Line: Probably yes on AI, but you’re asking the wrong question. Apple will integrate AI competently—they always do—but “successful market run” conflates product execution with stock returns. You’re not betting on Apple’s AI capability; you’re betting the market is underestimating them. That’s a harder case to make at a $3T+ valuation. (Medium confidence)
Question: I’m 60 with $200K to invest for retirement in 10-12 years. What’s the best way to maximize growth while minimizing taxes?
Bottom Line: You’re asking the wrong question, and no one can answer the right one without more information. “Best way to invest” requires knowing your full financial picture. You’re optimizing for “growth” when at 60 you should probably optimize for “risk-adjusted returns with sequence-of-returns protection.” The tax question is backwards—it depends on your current vs. future tax bracket, which you haven’t specified. (Medium confidence)
This example shows how be-critical protects you from getting (or giving yourself) bad advice—and includes a privacy-preserving template for asking about your own financial situation.
I kept having “conversations” with AI where every response was basically “That’s brilliant! Here’s how to make your already-perfect idea even better!”
And I get it. It feels good when AI validates your thinking. But if your AI assistant never pushes back on your ideas, you’re not using it to think sharply. You’re using it to feel better about decisions you’ve already made.
The goal isn’t to be mean. The goal is to surface truth. You can get validation anywhere. The value is in seeing what you can’t see and saying what others won’t say.
Be the critic you actually need, not the cheerleader you can easily find.
The adversarial prompting techniques in this skill were inspired by AI Prompt Hackers and their excellent article “10 Prompts That Force AI to Challenge Your Thinking.”
MIT