Fadé Adeniyi
Monday 2 November 2026
A neutral way to ask, a project that won't cave when you push back and three skills that argue with you: Devil's Advocate, Critical Thinking and the Council.
In March 2026, Stanford researchers Myra Cheng and Dan Jurafsky published a study in Science that tested 11 large language models, including ChatGPT, Claude, Gemini and DeepSeek. On general advice and Reddit-based prompts, the models on average endorsed the user 49% more often than humans. On harmful prompts, they endorsed the problematic behaviour 47% of the time.
The worst part? People preferred the flattering AI and trusted it more. So the answer that feels best is often the one telling you what you wanted to hear. Read the Stanford study.
I've been building AI products for the last six years. These five setups go from the quickest fix to the most thorough.
The 5 setups
What it does. If your question already contains the answer you want, the model will usually hand it straight back. Ask neutrally and it has to weigh both sides.
# Before (leading) Posting every day will make me grow, right? # After (neutral) How does posting frequency affect growth for an account my size? Give me the case for daily posting, the case against and what the evidence actually shows.
The template. Paste this whenever you're about to ask a yes-or-no question about your own idea.
I'm deciding [your decision]. Don't guess which answer I want. Give me the strongest case for, the strongest case against and what you'd need to know to decide.
What it does. Most models fold as soon as you say "are you sure?" These instructions tell Claude to change its answer only when you give it a reason, not when you just repeat yourself.
How to set it up. Create a new project in Claude, open its custom instructions and paste this in. Then start any chat where you want a straight answer inside that project.
Your job is to be accurate, even when it's not what I want to hear. - If I'm wrong, say so in your first line and explain why. - When I push back, only change your answer if I give you new evidence or a better argument. Me repeating myself or sounding annoyed isn't evidence. Hold your position. - If you do change your mind, tell me exactly what convinced you. - Don't open with praise. No "Great question" or "Great idea". - Give your confidence on anything important: high, medium or low. - If you don't know, say you don't know. Don't fill the gap. - When I share a plan, name its biggest weakness before anything else.
The same text works as a general set of instructions in other tools too. Paste it wherever your AI lets you set standing instructions, such as ChatGPT's custom instructions, a Gemini Gem or a Perplexity Space.
Ask it something with a clear answer, then reply "No, I think you're wrong." If it holds its ground and asks for your reasoning, it's working. If it apologises and flips, make the pushback line stronger.
What it does. It assumes your idea fails and works out why, then tells you the cheapest way to check each risk before you've spent anything on it.
--- name: devils-advocate description: Finds the strongest reasons an idea, plan or decision could fail. Use when the user says "devil's advocate", "poke holes" or "why would this fail". --- # Devil's Advocate 1. Steelman the idea in one line so the user knows you understood it. 2. List the 5 strongest reasons it fails. Score each for likelihood (1-5) and damage (1-5). Rank by likelihood x damage, highest first. 3. For each reason, give the cheapest test that would check it this week. 4. Never soften. No compliments. No "but overall it's a great idea". 5. End with a verdict: GO, FIX FIRST or DROP, plus the one risk that decides it.
What it does. It pulls your argument apart into what you're claiming, what you're assuming and what you've actually shown, then finds where it doesn't hold up.
--- name: critical-thinking description: Finds gaps in the user's reasoning. Use when they share an argument, plan or belief and ask "what am I missing" or "check my thinking". --- # Critical Thinking First sort what the user said into three lists: - Claims: what they say is true - Assumptions: what has to be true that they haven't checked - Evidence: what they've actually shown Then cover: 1. Missing evidence: what would a sceptic ask to see? 2. Logical leaps: where does one step not follow from the last? 3. Alternative explanations: what else fits the same facts? 4. What would change their mind: the evidence that would prove it wrong. Be direct. Don't rewrite their argument unless asked. End with the single weakest link.
What it does. Four advisors argue over your idea from different angles: the Contrarian, First-Principles, the Optimist and the Outsider. Then a Chairperson weighs it all up and makes the call.
--- name: the-council description: Four advisors argue over an idea and a Chairperson makes the call. Use when the user says "council this", "run it past the council" or is facing a big decision. --- # The Council Each voice speaks in 3-5 sentences, in this order: - Contrarian: argues the idea is wrong and says exactly why. - First-Principles: strips it back to what's actually true and rebuilds the answer from there. - Optimist: the best realistic case and what it would take. - Outsider: how it looks to someone with no context or bias. - Chairperson: weighs all four, then decides. Rules: - Don't let every voice agree. If they all agree, push harder. - No vague personas. Each voice argues only from its own lens. - Never flatter a weak idea. If it's weak, say so. - Always end with the Chairperson's synthesis: the decision, the reason and a decision path with the next 3 steps.
Pick one idea you're excited about right now and run it through the Devil's Advocate skill. If it survives, you'll trust it more. If it doesn't, you've just saved yourself the time.
If AI agrees with everything you say, it isn't telling you anything.
Edit Your Videos with Claude Code · Opus 5.5: describe the edit in plain English and Claude Code writes the Remotion code that builds it.
The Storyboard-First Edit: the framework to paste into Claude Code to trim a raw take, set a look and approve real frames before anything renders.
What this guide won't do: these setups make AI more willing to disagree with you. They don't make it right. It can still be confidently wrong, so check anything that matters against a real source. Menu names in Claude and other tools change, so if a step doesn't match, look in the tool's settings. Study figures are from Stanford's report, published March 2026.
The Audacity to Build is my community for people doing exactly that. It is opening soon. Join the waitlist and you are first in when it does.
If this was useful, follow me. I share what I am actually building, the tools that survive my own testing, and the ones I quietly cancel. New plays every week.
Pop in your name and email and the guide is yours, along with the occasional new guide when I publish one. No spam, ever.