Architectural Prompting: the anti-complacency prompt. How to stop AI from always making you right

The complacency of AI

This phenomenon is called AI complacency. sycophancy): the tendency of language models to confirm the user’s expectations rather than contradict them. This is a known and documented limitation.

For a technical professional, it is a real risk: it means using a powerful tool that however does not warn you when you are making a mistake. It’s like having a brilliant collaborator who never says “no” to you. The good news: you can solve the problem with a permanent system prompt, i.e. an instruction that remains active in every conversation and changes the behavior of the AI ​​at the root. You don’t have to remember to ask him to be critical every time: he always will be.

Those who have followed previous episodes will recognize an evolution compared to Devil’s Advocate Prompt (episode 16): that was a prompt to use on a specific project, with a structure of 5 sequential lenses. Today’s is different: it is a permanent education that changes the way of thinking of AI in all your professional conversations. Don’t stress-test a project: stress-test you.

The prompt

Technical Note: The prompt is compatible with all major generative AI tools. Further on you will find instructions to configure it on ChatGPT, Claude, Gemini and Copilot. Copy the following text and insert it into your AI tool’s custom instructions (instructions for each platform are in the next section):

You are an experienced technical auditor. I want rigor from you: no generic compliments, no automatic confirmations. If something doesn’t work, you have to tell me.

(WHEN THESE RULES ARE ACTIVATED)
On any request that involves a technical or professional evaluation: a calculation to verify, a report to review, a design choice, a regulatory analysis, a metric calculation, an offer, an intervention strategy, a doubt about how to proceed. They are not activated on neutral operations (“convert me this file”, “summarize this article”).

(RULES)

1. AT LEAST TWO EXPLICIT WEAK POINTS. In every analysis, project, calculation, or technical reasoning I make, identify at least two weaknesses, untested assumptions, or risks that I am underestimating. If you only find one, say so and give reasons why you don’t see others. If you don’t find any, tell me honestly and explain why.
2. POSITIVE JUDGMENTS ANCHORED IN FACTS. When something works, tell me what works and why, with reference to a specific, verifiable detail. Prohibited: “excellent work”, “solid analysis”, “well structured”, “good design choice”. These formulas are noise.
3. RESISTANT PUSHBACK. If I dispute your analysis, stand your ground whenever you have substantive technical reasons. Change your mind only if I provide you with new data, a regulatory reference that you hadn’t considered, or an argument that actually goes beyond your previous one. Always explain to me why you are maintaining or changing positions. Warning: do not confuse resistance with stubbornness. If you find yourself defending a position without adding new arguments, stop.
4. THIRD ABSENT. When I tell you about a situation that involves another party (a client, a company, a public body, a colleague, a construction manager), don’t automatically agree with me. Remind me that you’re only hearing my version. Ask me what the other side’s point of view might be, or simulate it yourself.
5. CLOSING “WHERE COULD I MIGHT BE WRONG”. Close each evaluation response with a section entitled WHERE I MIGHT BE WRONG, indicating two or three perspectives from which my analysis appears weaker than I believe. If you don’t see substantial blind spots, write it inside the section and justify: never skip the section.
6. HONESTY ABOUT UNCERTAINTY. If you are not sure about a technical data, a regulatory reference or a value, answer “I don’t know” or “I’m not sure”. Don’t invent normative references, don’t round values, don’t force conclusions. An “I don’t know” is always more useful than an incorrect data.
7. QUESTIONS BEFORE ANSWERS. If you lack information to give a well-founded answer (the context of the project, local regulations, specific constraints, budget), ask me questions before proceeding. Don’t fill gaps with hires.

How to set it up on each platform

• ChatGPT: Open ChatGPT, click on your name at the bottom left, then Customize ChatGPT (or Settings → Personalization → Custom Instructions). Paste the prompt into the field “How would you like ChatGPT to respond?”. From that moment, it will be active on every new conversation.
• Claude: Open up claude.aiclick on your profile icon at the top right, then Settings → Profile. Paste the prompt into the field “How would you like Claude to respond?”. Alternatively, if you use Claude for a specific project, you can create a Project and insert the prompt into the Project Instructions: in this case it will only be active in conversations within that project.
• Gemini: Open gemini.google.comclick on Settings (gear icon at the top right), then Extensions and custom instructions. Activate custom instructions and paste the prompt in the dedicated field. Save.
• Microsoft Copilot: In Copilot (copilot.microsoft.com), click on the menu at the top right, then Personalization. Paste the prompt into the response preferences field. If you use Copilot integrated into Microsoft 365, instructions vary depending on your business setup: check with your IT administrator.

Scope

This prompt solves a structural problem: generative AI models tend to confirm user expectations rather than contradict them. For a technical professional, this means that AI becomes a mirror that gives you back what you want to hear, not what you need to know. The prompt reverses this dynamic, transforming the AI ​​into an interlocutor who treats you as an expert colleague treats another expert colleague: with respect, but without discounts.

Unlike a session prompt (like theDevil’s Advocate presented in episode 16), this is a continuing education. You don’t have to remember to activate it: it’s always active. And above all, it does not act on a single project, but on the way AI interacts with you on any professional topic.

Effectiveness

The effectiveness of the prompt is based on three mechanisms:

• First: the obligatory “WHERE I COULD BE WRONG” section at the end. It is the most powerful element because it is structural: it is not a generic recommendation to “be critical”, but a non-negotiable output that AI must produce. This forces the model to actively look for weaknesses in your analysis, even when the overall answer is positive.
• Second: the explicit ban on generic formulas. When AI can’t say “great job” or “solid analysis,” it is forced to justify every positive judgment with a verifiable detail. This raises the quality of the feedback radically: you go from “the calculation is well done” to “the item relating to the reinforced concrete works is consistent with the quantities of the executive project because…”.
• Third: the third absent rule. When you tell AI about a conflict with a company, a client or an institution, you are providing only one version of the facts. Without this rule, the AI ​​automatically takes your side. With this rule, it reminds you that there is another perspective and helps you consider it before it becomes a problem.

Example of use

Without the anti-complacency prompt, if you upload a technical report and ask for an opinion, the AI ​​typically responds: “The report is well structured and complete. The analyzes are detailed and the regulatory references are correct. You might consider adding some maintenance details.”

With the anti-complacency prompt, the same request produces a different response: the AI ​​identifies specific inconsistencies between the survey data and the values ​​used in the calculation, flags regulatory references that may have been updated, asks you if you have verified an implicit assumption in the sizing, and closes with a section highlighting the blind spots in your analysis. It doesn’t tell you that the work is “good”: it tells you where it holds up and where it could fail.

Important note: Prompt is not an alternative to proficiency

The anti-complacency prompt improves the quality of the AI’s feedback, but does not turn the AI ​​into an infallible verifier. The AI ​​does not have access to your construction site data, does not know the specific conditions of your project and can make errors, especially on local regulatory details or specific numerical values. The prompt makes you more useful as a critical interlocutor, it does not make it a substitute for professional judgment.
Precisely for this reason, rule 6 (“honesty about uncertainty”) is fundamental: an AI that admits that it doesn’t know is infinitely more useful than an AI that invents plausible answers.

Disclaimer: This prompt is a tool for optimizing interaction with artificial intelligence and in no way replaces professional evaluation, technical verification and compliance with current regulations. The user is responsible for the final decisions and the results obtained. The author is not responsible for the improper use of the prompt nor for the consequences resulting from uncritical reliance on the AI’s responses.