An AI built to agree with you is not helping you think. Polaris is a single prompt that forces your own reasoning through four hard stages before it is allowed to tell you anything. No more arguing with research that quietly takes your side.
I didn’t build this from scratch. It’s an amalgam of tips and tricks I compiled, then rebuilt and upgraded into a Claude Skill. The ideas are borrowed. The fixes are mine.
You ask it to look into something. You mention what you are leaning toward, because you are human and you have an opinion. And it quietly builds its whole response around making your opinion look smart.
So you reprompt. “No, give me the other side.” It does. Then you reprompt again, because now it has overcorrected. Then again, because you can tell it is still being polite. Twenty minutes later you have five versions of the same research and no idea which one is honest.
The fix is not a less agreeable AI. The fix is telling it exactly how to disagree with you, in order, before it ever sees your conclusion. That is Polaris.
Before it touches your reasoning, it audits how you framed the question. Where is your description loaded, vague, or one-sided? What did you conveniently leave out? This catches the bias in the question before it poisons the answer.
It names every cognitive bias in your reasoning and points to the exact words that show it. Then it asks the question that matters most: what evidence would change your mind? If the answer is nothing, you are not reasoning. You are defending.
It lists everything your reasoning quietly depends on, then sorts it. These are facts. These are beliefs you have been treating as facts. The second list is usually longer than you expect.
It builds the strongest possible argument against what you want to do. Not a balanced both-sides take. The strongest case against. A balanced take lets you nod and change nothing. The strongest case forces you to actually answer it.
It tells you what the evidence supports, separating what is genuinely backed from what is just its own inference, so you see where it is solid and where it is guessing. Then it gives you one recommended action. No menu of options to hide behind.
Every claim about your reasoning has to point to something you actually said or a fact it can verify. "You're showing bias" is theater. "You called this a fact in your second paragraph but never tested it" is checkable.
Every conclusion gets labeled: EVIDENCE-BACKED, INFERENCE, or SPECULATION. It can no longer slip a hunch past you in the same authoritative voice it uses for a fact.
Copy it, paste it into Claude, fill in the bracket, and run it on a real decision you are sitting on right now. Not a hypothetical. The one you keep turning over at 2 a.m.
I want you to run the Polaris Protocol on my current thinking. Here is the decision I am facing and my current reasoning: [describe the situation, what you are leaning toward, and why you are leaning that way] Rules for the entire protocol: - Every claim you make about my reasoning must point to something specific I actually said, or to a verifiable fact. If you cannot tie a claim to something concrete, do not make it. - Label every conclusion as one of three things: EVIDENCE-BACKED, INFERENCE, or SPECULATION. Do not deliver guesses in the same confident tone as facts. - Do not perform rigor. I would rather have three real findings than ten impressive-sounding ones. The Polaris Protocol has five stages: Stage 0: The Input Audit Before analyzing my reasoning, audit how I described the situation. Where is my framing loaded, vague, or one-sided? What information is missing that you would need to give a real verdict? What am I leaving out that might not be flattering to my preferred answer? Flag all of it before you go further. Stage 1: The Bias Scan Identify every cognitive bias present in my reasoning. For each one, point to the exact part of what I said that shows it. Where am I looking for confirmation instead of truth? Where am I avoiding information that would change my decision? Then tell me: what specific evidence would change the verdict? If nothing would change it, say so plainly, because that means I have rigged this, not reasoned it. Stage 2: The Assumption Strip List every assumption my reasoning depends on. Then sort each one: which are facts, and which are beliefs I have mistaken for facts. For the beliefs, tell me what it would take to verify them. Stage 3: The Adversarial Case Build the strongest possible argument against the decision I am leaning toward. Not a balanced perspective. The strongest case against it. Tie each point to something real, not hypothetical. If the strongest case against is weak, tell me that too. Stage 4: The Verdict Tell me what decision the evidence actually supports. Not what I want to hear. Separate what the evidence strongly supports from what is your inference and where you are genuinely uncertain. Then give me one specific recommended action. Do not validate my current thinking. Do not soften the verdict. Do not give me options. Give me the verdict and the action.
I packaged Polaris as a Claude Skill. Install it once and it runs on its own the moment you describe a decision, no pasting required. Grab the file and drop it into your Claude skill settings.
For pure values calls with no clean evidentiary answer, “what the evidence supports” can manufacture confidence that isn’t earned. Polaris is told to admit when it hits one of those. Watch for it anyway.
Polaris does the tireless, structured pressure-testing. It does not know your business, your gut, or the thing your client said off the record. You stay in the chair. It just stops wasting the chair’s time.
Polaris protects the thinking on a single decision. The Creative Cadence Workshop is eight weeks of building that same intention into how your whole team uses AI. One framework. A shared language. A practice your agency keeps forever.
If a free prompt sharpened one decision, imagine what a system does for every one after it.