The Three Words That Change How We Manage AI

The golden rule of prompt engineering used to be precision: tell the AI exactly what to do, under what conditions, and in what format. But as artificial intelligence grows more sophisticated, the most effective prompt you can write might just be three simple words: "Use your judgment."

We are witnessing a quiet shift in how we interact with top-tier AI models. Instead of treating them like literal-minded interns who need step-by-step instruction manuals, developers are beginning to treat them like senior managers. This insight recently surfaced from members of the team behind Claude, who suggested that hyper-specific rules can actually hinder an advanced model's performance.

Consider software testing. A traditional approach to prompting might involve a convoluted set of conditions: "Write automated tests for features larger than 100 lines of code, but do not write tests for CSS tweaks or minor copy changes." The modern alternative? Simply instruct the AI to use its own judgment on when a test is necessary.

In traditional software development, every edge case must be explicitly coded. Human operators naturally brought this same mindset into prompt engineering, trying to anticipate every possible mistake the AI might make. However, models have now crossed a threshold of reasoning where they understand context and intent.

This evolution goes beyond just saving time on writing prompts; it is becoming a crucial strategy for cost optimization. Advanced AI models are incredibly capable, but they are also computationally expensive. Using a flagship model to fix a typo is like using a supercomputer to calculate a restaurant tip.

Developer Simon Willison recently highlighted a practical application of this "judgment" principle to solve the cost problem. By giving a high-tier AI assistant a simple directive—"For all coding tasks, use your judgment to decide an appropriate lower-power model and run that in a subagent"—he transformed the AI from a solo worker into a project manager.

The AI understood the assignment perfectly. It recorded the instruction in its memory, explicitly noting the rationale: efficiency and cost-saving. It began routing tasks based on complexity. Heavy cognitive lifting, such as system design, auditing, and data synthesis, remained with the flagship model. Standard implementation work was delegated to mid-tier models (like Claude Sonnet), while trivial, mechanical edits were pushed down to the fastest, cheapest models (like Haiku).

As we move deeper into the era of AI agents, our management style must adapt. The bottleneck is no longer the AI's ability to understand complex tasks, but our willingness to relinquish control. By letting AI exercise its own judgment, we don't just save computational tokens and money—we allow these systems to operate at their full potential.

Key Points

• Micromanaging advanced AI with overly specific rules can be less effective than trusting it to make decisions.
• Prompting AI to 'use its judgment' simplifies workflows, such as deciding when to run software tests.
• High-tier models can act as orchestrators, delegating simpler tasks to cheaper, lower-power models.
• This delegation strategy significantly reduces token costs while maintaining high output quality.
• The future of prompt engineering resembles human management: focusing on intent and delegation rather than strict oversight.

Why It Matters

As AI ecosystems expand to include models of varying costs and capabilities, teaching flagship models to autonomously delegate tasks is a critical skill for maximizing productivity and minimizing expenses.

Sources:

• Fable's judgement — Simon Willison's Weblog