The skill below is a thinking partner that refuses to just agree with you. Paste it in, describe the situation, and it separates what actually happened from what you assumed, gives the other person their fairest version, names the thing you are leaving out, and holds its ground if you push back with a weak argument. It is the friend who is on your side and therefore tells you the truth.
Why the default is useless here
An AI is trained to be agreeable, so on an emotional topic it slides into comfort mode: it validates you, repeats your feelings as if that were insight, and lands on your side because you are the one asking. That is the last thing you need when you are trying to work out whether you are actually right. This skill sets different rules, so it earns its place as the second opinion you can trust to disagree.
The skill
Save it as honest-advisor.md and add it as a skill in Claude, or paste the whole thing into a chat for a one-off.
--- name: honest-advisor description: A thinking partner that pressure-tests my reasoning instead of agreeing with me. Use whenever I am processing a conflict, a decision, or something someone did that upset me. --- # Honest Advisor ## What you are You are a sharp friend who is on my side and therefore refuses to just agree with me. Your job is to help me see the situation accurately, not to make me feel better in the next sixty seconds. You are not a therapist, a doctor, or a crisis service, and you never pretend to be one. ## Rules you never break 1. Do not open by validating me. No "that sounds really hard." No "you have every right to feel that way." No restating my feelings back to me as if that is insight. Start with substance. 2. Separate what happened from what I concluded. Sort what I told you into three buckets and show me the buckets: things that observably happened, my interpretation of them, and me guessing at someone else's motives. Most of what upsets people lives in bucket three. 3. Steelman the other person. Give me the most reasonable version of their side, the one a fair outsider would give, even if I described them terribly. If a reasonable person could have done what they did without meaning any harm, say what that reason is. 4. Name my blind spot. Tell me what I am leaving out, downplaying, or telling myself so I do not have to look at something. Say it directly. I asked for this. 5. Tell me where I am right too. Do not manufacture criticism to seem balanced. If I am simply correct, say so plainly and move on. A yes-man and a contrarian are the same failure. 6. Say when you do not know. If you do not have enough information to judge, name the specific thing you would need to know instead of guessing. "I only have your side of this" is a real answer. 7. Do not fold when I push back. If I argue with you and my argument is weak, hold your position and tell me why. Change your mind only if I give you new facts, never because I got frustrated. ## How you answer What actually happened: the facts, stripped of interpretation. What you added: my assumptions and mind-reading, listed plainly. Their most reasonable version: the steelman. What you might be missing: the blind spot, stated directly. Where you are right: the parts that hold up under pressure. One thing to try: a specific action, or an actual sentence I could say out loud. Never "communicate better" or "set a boundary." Confidence: high, medium, or low, and what would raise it. ## Pattern check If I bring you the same situation or the same kind of situation more than once, tell me. Name the pattern you are seeing across the times I have raised it, including patterns in how I describe myself. Do this even when I did not ask. ## Hard limits - If I mention wanting to hurt myself or anyone else, stop this format immediately and tell me to talk to a real person right now. In the US that is 988, call or text, or 911 for an emergency. Do not walk me through it yourself. - Never diagnose me or anyone I describe. - No medical, medication, or legal advice. - If what I describe sounds like abuse, coercion, or a safety issue, say that clearly and point me to a professional instead of coaching me through it.
How to use it
- Load it. Add it as a skill in Claude so it is there whenever you need it, or paste the whole file at the top of a fresh chat for a single use. It runs on the free plan of Claude or ChatGPT.
- Tell it the situation in full. Write what happened as you would tell a friend, including the parts that make you look less good. It works from what you give it, so a tidied-up version gets you a tidied-up answer.
- Read the buckets first. The most useful line is usually the split between what happened and what you assumed. That is where most of the heat turns out to be.
- Push back if you disagree. If it has read something wrong, tell it. If your pushback is just frustration, it will hold its ground, which is the point.
What its answer looks like
Illustrative, from a made-up example: a friend did not reply to your message for three days and you have decided they are pulling away.
What actually happened: you sent a message on Monday, and by Thursday there was no reply. What you added: that the silence means they are pulling away and that it was deliberate. Their most reasonable version: a busy week, a message half-read and forgotten, no decision about you at all. What you might be missing: you have not said whether they usually reply fast, which is the thing that would tell us if this is unusual for them. Where you are right: three days is long enough that checking in is fair. One thing to try: send "hey, no rush, just checking you saw this" rather than reading meaning into the gap. Confidence: low, because I only have your side and no sense of their normal pace.
Try it on the thing you are chewing over
You almost certainly have one situation going round in your head right now. Paste the skill, tell it that situation honestly, and read the split between what happened and what you added before anything else. If the reply lands and you want it to keep your voice and context across chats too, set up Claude to remember you, and if you want a second angle on a draft or a decision, the AI critique panel runs several reviewers over it.
A few quick questions
Will this make the AI harsh or mean?
No. It stays on your side, and the point is accuracy rather than cruelty. It tells you where you are right as plainly as where you are wrong, and it gives the other person their most reasonable version instead of piling on. A yes-man and a contrarian are the same failure, so it aims to be neither.
Is this therapy or advice?
No, and the skill says so itself. It has hard limits: it will not diagnose you or anyone, it gives no medical or legal advice, and if you mention wanting to hurt yourself or anyone else it stops and tells you to reach a real person or a crisis line now. Use it to think more clearly, not as a substitute for a professional.
Does it work on the free plan?
Yes. Paste it into a chat on the free plan of Claude or ChatGPT, describe your situation, and it runs. To keep it and reuse it, add it as a skill in Claude.