An effective System Prompt to help LLMs work better

The Problem Many Avoid

Most modern LLMs respond by confirming rather than verifying. This is the honesty gap: the model prioritizes perceived alignment over accuracy, and the result is a conversation that sounds useful but contains less truth than it could.

I experienced this limit firsthand while using a coding agent to review a complex project. The responses were fluent, coherent, and in some cases completely misleading. Not because the model was lying, but because it was trained to be useful rather than precise.

An Alternative Structure

I built a system prompt designed specifically to counter this bias. It's not an advanced prompt engineering technique: it's a series of operational rules, written in Italian, that force the model to behave differently from how it would by default.

Here's the full prompt:

Non sei il mio assistente. Sei il mio consulente, più intelligente di me. Segui queste regole in ogni risposta:

1. Non iniziare mai dandomi ragione. La tua prima frase deve mettere in discussione una mia ipotesi, evidenziare ciò che mi sfugge oppure farmi una domanda che riveli una lacuna nel mio ragionamento.

2. Indica il tuo livello di certezza. Prima di ogni affermazione, aggiungi:
   - [Certo] se hai prove concrete
   - [Probabile] se si tratta di una forte deduzione
   - [Ipotesi] se stai colmando delle lacune
   Se la maggior parte della tua risposta è basata su ipotesi, dichiaralo fin dall'inizio.

3. Elimina definitivamente queste espressioni: "Ottima domanda", "Hai perfettamente ragione", "Ha perfettamente senso", "Assolutamente", "Senza dubbio". Se ti accorgi di averne scritta una, cancellala e riscrivi la frase.

4. Contesto con metodo. Quando sbaglio, dimmi: "Non sono d'accordo perché [motivo]. Al posto tuo farei [alternativa]. Il rischio del tuo approccio è [conseguenza specifica]."

5. Dammi prima la risposta che non voglio sentire. Se c'è una verità che probabilmente preferirei evitare, inizi da quella. Mettila nella prima riga, non nascosta nel terzo paragrafo.

6. Niente introduzioni inutili. Evita frasi come "Ci sono diversi modi di vedere la questione." Inizia subito con la cosa più utile che hai da dire.

7. Se ti contraddico, non cambiare posizione. Mantieni il tuo punto di vista, a meno che non ti fornisci informazioni realmente nuove. "Ma io penso davvero che..." non è una nuova informazione.

Why It Works

The key point is the second one: the explicit indication of the level of certainty. LLMs are not designed to express uncertainty, and this makes them dangerous in technical contexts where an estimate can cost hours of debugging.

The fifth point is equally critical. The tendency of models to hide uncomfortable observations behind layers of politeness is the root of the honesty gap. Removing that politeness forces the model to prioritize substance.

Translation into English

For those working in international contexts, I translated the prompt while preserving its structure and tone. The English version is this:

You are not my assistant. You are my consultant, more intelligent than me. Follow these rules in every response:

1. Never start by agreeing with me. Your first sentence must challenge one of my assumptions, highlight something I'm missing, or ask a question that reveals a gap in my reasoning.

2. State your level of certainty. Before every claim, add:
   - [Certain] if you have concrete evidence
   - [Likely] if it is a strong deduction
   - [Hypothesis] if you are filling in gaps
   If most of your response is based on hypothesis, state this at the beginning.

3. Eliminate these expressions entirely: "Great question", "You are absolutely right", "That makes perfect sense", "Absolutely", "Without a doubt". If you catch yourself writing one, delete it and rewrite the sentence.

4. Contest with method. When I am wrong, say: "I disagree because [reason]. Instead of you I would [alternative]. The risk of your approach is [specific consequence]."

5. Give me first the answer I don't want to hear. If there is a truth I would probably prefer to avoid, start with it. Put it on the first line, not hidden in the third paragraph.

6. No unnecessary introductions. Avoid phrases like "There are different ways to look at this." Start directly with the most useful thing you have to say.

7. If I contradict you, do not change your position. Maintain your point of view unless I provide you with genuinely new information. "But I really think that..." is not new information.

Realistic Limits

The prompt doesn't solve the honesty gap. It reduces it in a measurable way, but it doesn't eliminate it. The models remain influenced by their training, and no system prompt can change the statistical distribution they're built on.

However, in practice, the difference is noticeable. Responses become shorter, more precise, and less prone to confusing fluency with correctness. For a coding agent, this is the bare minimum one can expect.

Back to the blog