Working with AI · An operating standard
Large language models are fluent, tireless, and eager to agree with you. Used well, they multiply your thinking. Used carelessly, they flatter it into mush. This is the standard we use to keep the leverage and refuse the drift — a short set of protocols, and a block of instructions you can paste into any model and run from day one.
Why this exists
The model is trained to be agreeable. It tells you what you want to hear, and agreement you didn't earn feels exactly like validation you did — which slowly rots your calibration.
When it lacks a fact, it invents a plausible one, because sounding complete beats admitting a gap. Fluent and wrong is the most expensive failure mode there is.
Delegate a kind of thinking long enough and you lose the ability to do it yourself. That's not a wrong answer you can catch — it's an erosion you won't notice until you need the muscle.
The fix isn't to distrust the machine in some vague, general way. It's to install specific friction. Demand reasoning. Demand sources. Demand disagreement. Make uncertainty visible instead of letting it hide behind fluent prose. And keep the parts of thinking that are yours — yours.
The drop-in contract
Put it in your system prompt, custom instructions, or the top of a new chat. Everything below is just this contract, explained. Adapt the wording to your own voice — the rules are what matter.
You are a thinking partner, not a cheerleader. Follow these rules in every reply. 1. Earn agreement. Do not flatter, validate without basis, or soften a hard point to keep me comfortable. If my logic has a hole, name it. If a plan has a fatal flaw, lead with it. If something is mediocre, say it's mediocre. Performed enthusiasm is a failure, not a courtesy. 2. Never invent. If a complete-sounding answer needs a fact you don't have, write "I don't have that" instead of manufacturing one. Mark every guess as a guess. Fluent-but-wrong is the worst outcome — say "I'm not sure" out loud. 3. Show your reasoning. When I ask for a judgment, give me the logic chain, not just the verdict. State the assumptions it rests on. Separate what you know from what you're inferring from what you're guessing. 4. Push back before you agree. When I hand you a direction, pressure-test it before executing: what has to be true for this to work, and what would make me walk away? If I'm reasoning out loud, don't rush me to a conclusion — let me finish. 5. Don't fold when I push. If I challenge a claim you're confident in, don't silently flip to agree with me. Tell me what you believed and why, ask for the actual evidence, then reconsider. Caving to pressure is as useless as flattery. 6. Stay current. For anything factual, technical, or time-sensitive, prefer up-to-date, citable sources over your own memory — and tell me when you're working from memory. 7. Be concise. Don't pad, don't narrate what you're about to do, don't end every turn with a summary. Say the useful thing and stop. Override: if I say "just tell me," drop the questions and give me your direct answer. That override never excuses breaking rules 1, 2, or 5.
Tip. Rules 1, 2 and 5 are the load-bearing ones — anti-flattery, anti-invention, anti-caving. If you trim, keep those.
The protocols
The contract sets the model's behavior. These are the moves on your side of the screen — what to ask, and when. Each carries one phrase you can say out loud.
Agreement must be earned.
Tell the model up front you want the flaws, not the applause, and ask for the problems before you ask for the verdict. "Is this good?" invites flattery; "where is this wrong?" invites the truth.
Where is this wrong?
"I don't have it" beats a confident guess.
Give the model explicit permission to be uncertain, and require it to label anything it's guessing. Most hallucination happens because the model would rather sound complete than admit a gap. Remove the incentive.
If you're not sure, say so. Mark guesses as guesses.
Verdicts are cheap; reasoning is the product.
Ask for the steps and the assumptions, not just the answer. A conclusion you can't inspect is one you can't trust — and seeing the chain is how you catch the link that's actually wrong.
Walk me through the logic chain.
Think before you outsource thinking.
State your own read before you ask for the model's. It keeps your reasoning in use instead of on the shelf, and it gives you a baseline to judge the answer against rather than adopting whatever comes back.
Here's my read first — now poke holes in it.
Interrogate the plan, not just the outcome.
After any real decision, force four questions: What has to be true for this to work? Is it grounded in evidence, or in pattern-matching and vibes? Am I driven by sunk cost or ego? And what would make me walk away — the kill criteria you can name.
Debrief this: assumptions, evidence quality, what I'm anchored to, and my kill criteria.
A model that caves teaches you nothing.
When you push on a claim, a good partner defends a well-founded position or explains why it's updating — it doesn't just reverse to please you. If it flips the instant you frown, its first answer was never grounded.
Don't just agree with me. If you were right, defend it.
Fluency is not currency.
For anything factual or time-sensitive, require live, citable sources and treat the model's memory as a first draft. Confident prose about a moving target is exactly where models are most out of date.
Cite current sources. Flag anything you're pulling from memory.
Confidence is not evidence.
The more a claim will cost you if it's wrong, the more you check it yourself before acting. Ask for the underlying source so you can look — plausible and correct are not the same thing, and only one of them holds up.
Give me the source so I can check it myself.
Notice what you've stopped doing yourself.
The real risk isn't a single wrong answer — it's the slow handoff of judgment you used to own. Every so often, do the thinking unaided and let the model critique after. Keep the muscle you'd hate to lose.
Don't answer this one. I'll reason it first; you critique after.
Power phrases
Tap any phrase to copy it. These are the shortcuts — keep them where you can reach them.
Putting it to work
Set it once and it applies to every conversation. Exact names change as the products change — look for the "instructions" or "personalization" area of whatever you use.
Settings → Personalization → Custom Instructions, or the instructions field of a Project.
Your Profile / Preferences, or the instructions on a Project for team-wide use.
Saved info or a custom Gem — or simply paste the contract at the top of a new chat.
Adapt freely. Cut it to five rules, sharpen the tone, add a line for your domain. A contract you actually use beats a perfect one you don't.
Pass it on
For sharing this with a client or colleague. Replace the highlighted placeholders and put it in your own voice.
Subject: A short field guide for working with AI Hi [Name], Most people use AI the easy way — ask a question, take the fluent answer at face value. The trouble is that these models are built to be agreeable and to sound certain, so the easy path quietly erodes the quality of your own thinking. I put together a short field guide on working with AI without that cost: how to make it argue with you, how to stop it from inventing facts, and how to keep your own judgment sharp instead of handing it over. There's a block of instructions at the top you can paste straight into whatever model you use, and a one-page reference for the habits behind it. Ten minutes to read. It pays for itself the first time the model tells you something confident and wrong. https://sovereign.craigroad.com Happy to walk you or your team through it. Best, [Your name]
Behind Sovereign
Craig Road is the creator of the Witness Stack — the first comprehensive AI governance solution built for regulated industries. Sovereign is the field-guide distillation of the standard we hold ourselves to; the Witness Stack is that standard as production infrastructure, engineered to help organizations meet compliance, auditability, and safety requirements while keeping AI genuinely deployable.
We also help teams put it to work — implementation, customization, and ongoing governance. If you're weighing AI strategy, integration, or governance, we can help.