Boris Cherny Says Give Claude Goals, Not Instructions

Stop micromanaging AI? Claude Code's creator has a different approach

Boris Cherny, one of the creators of Claude Code at Anthropic, is highlighting a simpler approach to prompting AI agents: tell the model what you want to achieve instead of explaining every step it should take.

In a recent post discussed by The Times of India, Cherny said users should approach Claude more like a colleague. Give it a clear goal and enough context, rather than trying to control every individual action. He also noted that modern models can determine much of the execution themselves once they understand the objective. The Times of India

This represents an important change in prompting philosophy.

The traditional approach often looks like:

20 detailed instructions → Step 1 → Step 2 → Step 3 → Step 4 → Final result

The agentic approach is closer to:

Goal + Context + Constraints → AI plans → AI executes → AI verifies

For example, instead of telling a coding agent exactly which files to open, which functions to modify and which commands to run, a developer could explain the problem, provide the relevant context and define what a successful result should look like.

Claude can then determine the best sequence of actions.

Anthropic's own research supports part of this broader shift. An analysis of around 400,000 Claude Code sessions found that people generally make the major planning decisions—what to do—while Claude makes more of the execution decisions—how to do it. The research also found that users with greater domain expertise tend to get more work done per instruction. Anthropic

However, this does not mean that longer prompts are always bad.

Complex tasks still need relevant context, examples, technical requirements and constraints. Anthropic's Claude Code guidance recommends being precise about expected behavior and providing the information the agent actually needs. Anthropic Resources

The bigger lesson is that prompting is moving from instruction writing toward delegation.

Instead of thinking, “How do I tell AI every step?”, users increasingly need to think, “What goal should I give the AI, what context does it need, and what boundaries should I set?”

For creators and developers, this could make AI agents much easier to use for larger tasks such as coding, research, content creation and data analysis