--- layout: default title: be-critical description: A Claude Code skill that transforms AI from yes-man to adversarial critic --- # be-critical **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](https://github.com/pem725/be-critical) | [Download v0.3](https://github.com/pem725/be-critical/releases/tag/v0.3) ---

The Problem

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. ### The Solution This skill forces Claude to do the opposite: - Build strong counterarguments - Identify unstated assumptions - Detect logical fallacies and cognitive biases - Say what others won't say **If reading the response doesn't make you uncomfortable, it wasn't critical enough.** ---

Installation Guide

### Method 1: Claude.ai Settings (Easiest) The simplest way to install this skill is through the Claude.ai web interface: 1. Download [be-critical-v0.3.zip](https://github.com/pem725/be-critical/releases/download/v0.3/be-critical-v0.3.zip) 2. Go to [claude.ai/settings/capabilities](https://claude.ai/settings/capabilities) 3. Scroll to the **Skills** section 4. Either: - Click **Add Skill** and select the zip file, or - **Drag and drop** the zip file directly onto the page 5. The skill is now available in your Claude conversations **Using it:** Just type `/be-critical` in any Claude conversation and the skill will activate. --- ### Method 2: Claude Code CLI If you use Claude Code (Anthropic's command-line interface), you can install the skill locally: #### Prerequisites You need Claude Code installed. If you don't have it: ```bash npm install -g @anthropic-ai/claude-code ``` > **Don't have npm?** Download Node.js from [nodejs.org](https://nodejs.org/) first. #### Install the Skill **On Mac/Linux:** ```bash mkdir -p ~/.claude/skills cd ~/.claude/skills git clone https://github.com/pem725/be-critical.git ``` **On Windows (PowerShell):** ```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](https://git-scm.com/downloads) #### Using the Skill 1. Start Claude Code by running `claude` in your terminal 2. Type `/be-critical` and press Enter 3. Claude will ask what kind of critique you need 4. Present your idea, decision, or question 5. Get challenged! --- ### Method 3: Manual Download If you prefer not to use git: 1. Go to the [Releases page](https://github.com/pem725/be-critical/releases) 2. Download `be-critical-v0.3.zip` 3. Extract the zip file 4. Either: - **For Claude.ai:** Upload `Skill.md` at [claude.ai/settings/capabilities](https://claude.ai/settings/capabilities) - **For Claude Code:** Move the folder to `~/.claude/skills/be-critical` --- ### Troubleshooting | 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](https://git-scm.com/downloads) | | "command not found: npm" | Node.js isn't installed. Download from [nodejs.org](https://nodejs.org/) | | "/be-critical" not recognized in Claude Code | Restart Claude Code after installing the skill | ---

How It Works

### 1. One Question to Focus 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 | ### 2. BLUF First 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. ### 3. Multi-Layer Analysis The skill applies 3-5 techniques in sequence, where each layer reveals what previous ones missed: - **Layer 1** reveals you're assuming a problem exists - **Layer 2** shows you're optimizing for feeling productive over validating demand - **Layer 3** identifies this is the third time you've done this pattern - **Layer 4** exposes you're afraid of sales conversations so you hide in building By Layer 4, you've gone from "some risks" to "here's the psychological pattern that keeps sabotaging you." ---

What It Detects

### 40+ Logical Fallacies | 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 | ### 20+ Cognitive Biases | 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 | ### Analysis Techniques | 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 | ---

Two-Stage Manuscript Review Protocol

For peer review of academic manuscripts, `be-critical` runs a specialized two-stage workflow defined in [`manuscript_review_protocol.md`](https://github.com/pem725/be-critical/blob/main/manuscript_review_protocol.md). ### Pre-flight: Journal Quality Assessment Before reviewing, the protocol characterizes the target venue — tier, indexing, audience, typical methodological standard. **Critique severity is calibrated to the journal's published bar, not to top general-science venues.** A solid mid-tier specialty journal does not need to clear a Nature standard. ### Stage 1 — Summary, Logic Model, Questions Read the paper, then produce a single markdown artifact containing: - Plain-language summary of the contribution - Verbatim research questions and hypotheses - A **Mermaid logic model** showing theory → predictions → design → measures → analysis → findings → claims, with explicit "inferential leap" branches where the chain skips a step - **Evidence map** — each claim mapped to the evidence cited and a strength rating - **5–10 categorized provocations** organized as logical / evidential / framing / structural problems - **3–6 targeted questions** for the human reviewer (verdict instinct, severity dials, field-specific calls Claude cannot make) The protocol then **stops** and waits for the reviewer's answers. Stage 1 is neutral-to-mildly-skeptical inventory, not a verdict. ### Stage 2 — Review in the Reviewer's Voice Using the reviewer's answers to calibrate severity and priority, produce a review in a two-letter format: - **To Editor:** short — compresses the paper's contribution into two sentences, conveys the recommendation through tone, flags reviewer calibration if needed - **To Authors:** opens with what is genuinely strong, then organizes critique under 3–5 named-header sections with numbered points (location pointer + why it matters + recommended literature where useful), closes constructively See the [worked example](https://github.com/pem725/be-critical/blob/main/examples/manuscript-review-stage2.md) for the format applied to a real review (redacted for peer-review confidentiality). ---

Examples

### Example 1: Solar Panel Decision > **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)* [Read full analysis →](https://github.com/pem725/be-critical/blob/main/examples/solar-panels.md) --- ### Example 2: Electric Vehicles vs ICE > **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)* [Read full analysis →](https://github.com/pem725/be-critical/blob/main/examples/ev-vs-ice.md) --- ### Example 3: Apple AI Investment > **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)* [Read full analysis →](https://github.com/pem725/be-critical/blob/main/examples/apple-ai-investment.md) --- ### Example 4: Retirement Investment Strategy > **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. [Read full analysis →](https://github.com/pem725/be-critical/blob/main/examples/retirement-investment.md) --- ## Why I Built This 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. ---

Credits

The adversarial prompting techniques in this skill were inspired by [AI Prompt Hackers](https://www.aiprompthackers.com/p/10-prompts-that-force-ai-to-challenge) and their excellent article "10 Prompts That Force AI to Challenge Your Thinking." --- ## License MIT ---

be-critical

Install this skill

Comments