Federico Donatone’s AI Prompting Strategy: Eight Rules for Better Claude Output

In a recent LinkedIn post, Federico Donatone discusses a novel approach to improving the reliability and output of AI language models, specifically focusing on Anthropic’s Claude. Donatone details a method involving a custom instruction file, dubbed CLAUDE.md, designed to guide the AI’s behavior and prevent common errors.

The core of Donatone’s strategy lies in establishing clear, actionable rules that Claude must adhere to before processing requests. This proactive measure aims to mitigate issues where the AI might prematurely declare tasks complete or fail to provide necessary evidence. Donatone highlights the frustration that can arise from such AI missteps, stating:

“Claude kept saying the work was done. It wasn’t.”

To address this, Donatone developed a set of eight rules embedded within the CLAUDE.md file. This file, according to Donatone, is intended to be pasted into the AI’s settings once, ensuring that Claude reads and internalizes these directives before undertaking any task.

Establishing Ground Rules for AI Assistants

Federico Donatone emphasizes the importance of verifiable progress and self-correction in AI interactions. His rules aim to instill a more rigorous workflow within the language model. As Donatone notes, the principle of continuous improvement is central to his method:

“Every correction becomes a permanent rule.”

This rule suggests a learning mechanism where identified flaws are not just fixed for a single instance but are codified to prevent recurrence. Donatone also stresses the need for clear communication and confirmation, advocating for explicit requests for information and confirmation before executing critical actions.

Key Directives for AI Interaction

The CLAUDE.md file, as outlined by Donatone, includes specific instructions designed to optimize the AI’s performance and user experience. These directives cover a range of scenarios, from handling reversible tasks to managing interruptions.

• Proof of Completion: Donatone’s first rule insists that an AI should never claim a task is finished without providing evidence.
• Self-Disproving Work: A key tenet is for the AI to actively try to find flaws in its own work before submission, promoting a higher standard of output.
• Confirmation Protocols: The strategy mandates seeking confirmation before irreversible actions are taken, safeguarding against unintended consequences.
• Independent Action on Reversible Tasks: Donatone suggests that reversible tasks can be performed without explicit permission, streamlining workflow.
• Clear Communication: A crucial rule requires users to state upfront if they need anything from the AI, preventing ambiguity.
• Learning and Tool Creation: Donatone encourages the AI to transform learned information into reusable tools, fostering efficiency.
• Minimizing Interruptions: The rules also address the user experience, with a directive to “Work without interrupting my screen.”

Donatone’s approach, detailed in his post, is a practical application of prompt engineering, aiming to create a more robust and predictable interaction with advanced AI models. By codifying these rules, he seeks to transform the AI from a sometimes-unreliable assistant into a more dependable collaborator.

According to Federico Donatone, the implementation is straightforward:

“It’s a CLAUDE.md file. Paste it into your settings once. Claude reads it before taking action.”

This method offers a compelling example of how users can actively shape AI behavior to better meet specific needs, turning potential AI shortcomings into opportunities for enhanced performance through structured guidance.