The Problem: A Polished Text Isn't a Correct Text
Generative AIs are designed to please. They respond in a confident, well-structured tone, even when they are wrong. Faced with a clean result, the temptation is to accept it. Researchers call this automation bias: the tendency to overestimate the reliability of an automated system, notably described by Raja Parasuraman and Dietrich Manzey in 2010. In business, the phenomenon was named "workslop" in 2025: AI-generated content that looks like finished work but shifts the effort onto the colleague who receives it.
The Five Steps
- Score. Have the response rated from 1 to 10.
- Criticise. Have it list every weakness.
- Optimize. Have it build a better version.
- Retry. Answer any questions raised, then relaunch.
- Execute. Adjust and apply the final touches yourself, without the AI.
The SCORE© Prompt Template
I asked an AI the following:
{original prompt}
I received the following result:
{paste result}
SCORE© THIS RESPONSE OUT OF 10.
BE VERY DEMANDING. SHOW THE WEAKNESSES.
PROVIDE A BETTER VERSION OR ASK QUESTIONS IF YOU LACK INFORMATION.
SCORE©: proprietary Revenu IA method, © Mustapha Benkalfate. Reproduction of the prompt template is authorized with source attribution.
Three Rules for Use
- Always in a new chat, without revealing the source. The AI judges the result, not the effort of the previous conversation.
- Read the list of weaknesses rather than the score. The scores are generous: "a 2 becomes a 4," observes Mustapha Benkalfate. What matters are the identified flaws.
- Iterate. For a business proposal or high-visibility content, several passes are useful.
Why the Last Step is Done Without the AI
Applying the final touches yourself confirms that you''ve understood what was wrong, rather than just assuming it. This is what distinguishes a tool-assisted review from complete delegation. This final step aligns with the Methodology of Doubt©: for anything that involves your credibility or the company''s, the AI can review and challenge, but the final delivery is yours.
A Variation for Data-Driven Reports
For a report or analysis, Mustapha Benkalfate suggests a simple addition: ask the same question in three separate chats. If the results diverge significantly, they cannot be trusted; if they converge, you can build upon them. He also recommends asking for a confidence score and sources, then reading the best ones yourself.
Choosing the Right Method
| Your Situation | Method | What It Does |
|---|---|---|
| You know exactly what you want | SCOPE© | Frames the request along five axes |
| Your need is still unclear | PROBE© | The AI asks you questions before generating |
| The task is too complex for a single request | STEP© | Breaks it down into blocks, tasks, steps, processes |
| You have a model result to replicate | SEAT© | Finds the prompt that produced it |
| You need to verify an AI's output | SCORE© | Has the result rated and critiqued by another AI |