Yonathan Cohen on the Evolving AI Prompting Landscape: Judgment Over Hard Rules

In a recent LinkedIn post, Yonathan Cohen discusses a significant shift in how we interact with and instruct artificial intelligence models, particularly in light of Anthropic’s recent changes to Claude’s system prompt. Cohen highlights that the era of heavily prescriptive, rigid instructions is giving way to a more nuanced approach centered on clear goals and model judgment.

Cohen points out the outdated nature of previous prompting strategies. He notes:

“For two years, everyone stacked rules: ‘never do this, always do that, use this exact format.’
Anthropic just removed most of theirs from Claude’s own system prompt, and published why.”

This move by Anthropic, as detailed by Cohen, signals a fundamental change in AI development and deployment. The core message is that newer, more capable AI models do not require the extensive, hard-coded constraints that were necessary for earlier iterations.

The Shift from Hard Rules to Judgment

A central theme in Yonathan Cohen’s analysis is the move away from rigid, absolute commands towards a framework that leverages the AI’s enhanced judgment capabilities. Cohen argues that the old methods of instruction are now counterproductive.

He elaborates on this point, stating:

“the new models don’t need ‘never write comments’. They need to know the goal.”

According to Cohen, this means that prompt engineers and users must adapt their strategies. Instead of exhaustively listing prohibitions and mandatory actions, the focus should be on defining the desired outcome or objective. This allows the more advanced AI models to utilize their improved understanding and reasoning abilities to achieve the goal effectively, rather than being constrained by a lengthy list of “dos and don’ts” that may no longer be relevant or even applicable to their advanced architecture.

Rethinking Instruction Structure: From Monoliths to Modularity

Yonathan Cohen also addresses the structural changes needed in how AI instructions are organized. He suggests that the practice of consolidating all instructions into a single, large document is becoming obsolete.

Cohen advocates for a more modular approach:

“stop front-loading everything into one giant instruction file.
Small files, loaded when the task needs them. (Yes: your AI setup should look like folders.)”

This suggests a paradigm shift towards dynamic instruction loading, where specific sets of instructions are activated based on the immediate task requirements. This approach not only streamlines the primary interaction but also allows for greater flexibility and efficiency, much like organizing files into distinct folders for easy access and management.

The Problem with Outdated Examples

Furthermore, Cohen highlights a critical issue with using old examples to guide newer AI models. He posits that examples meticulously crafted to constrain weaker models can inadvertently limit the potential of more powerful ones.

Cohen explains:

“examples written to constrain a weaker model box in a stronger one.
Show the outcome you want, not the exact steps you used to force.”

This implies that the focus of examples should evolve. Instead of demonstrating the precise, step-by-step process that was necessary to achieve a result with older AI, users should now present the desired end-state. This encourages the AI to find its own, potentially more efficient or creative, path to the solution, leveraging its advanced capabilities.

Auditing and Adapting AI Playbooks

To facilitate this transition, Yonathan Cohen points to new tools and methodologies, such as the `/doctor` command mentioned in relation to Anthropic’s models. This command serves as an audit tool, identifying instructions that are relics of older AI paradigms.

Cohen warns:

“run it. You’ll be surprised how much of your setup is archaeology.”

He urges readers to apply these insights to their own AI strategies and team playbooks. If these playbooks are heavily laden with “always/never” lists drafted months ago, they risk hindering rather than helping the AI’s performance. Cohen concludes that while weak models required rules, strong models thrive on clearly defined goals. This fundamental difference necessitates a complete re-evaluation of how we brief and manage our AI interactions.