> ## Content Index
> Fetch the complete content index at: https://wire.fourthweb.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# Stop Writing Detailed Prompts—Claude's Creator Says You're Making It Dumber
- URL: https://wire.fourthweb.ai/stop-writing-detailed-prompts-claudes-creator-says-youre-making-it-dumber/
- Published: 2026-07-27T22:08:02.000Z
- Updated: 2026-07-27T22:32:25.000Z
- Description: The better AI gets, the worse your old prompting habits become. Boris Cherny, creator of Anthropic's Claude Code, says users underestimate newer models by giving overly detailed step-by-step instructions — the equivalent of micromanaging a senior engineer
- Author: Travis Wright
- Tags: AI Agent Economy, AI Agents, DeFi, Anthropic

**The better AI gets, the worse your old prompting habits become.**

### The Summary

- [Boris Cherny, creator of Anthropic's Claude Code, says users underestimate newer models by giving overly detailed step-by-step instructions](https://www.businessinsider.com/anthropic-claude-code-prompting-tips-boris-cherny-micromanaging-ai-2026-7?ref=wire.fourthweb.ai) — the equivalent of micromanaging a senior engineer
- Better approach: Define the outcome, set guardrails, let the model choose its own path
- Google Brain cofounder Andrew Ng calls this "lazy prompting" — a feature of capable models, not a bug
- What worked six months ago actively hurts performance today

### The Signal

[Boris Cherny stood in front of Y Combinator's crowd](https://www.businessinsider.com/anthropic-claude-code-prompting-tips-boris-cherny-micromanaging-ai-2026-7?ref=wire.fourthweb.ai) and told them they're managing AI like it's still 2023\. The step-by-step instructions, the hand-holding, the "do this, then this, then this" prompts — all of it is slowing down models that can now handle abstract goals.

This isn't about being lazy. It's about recognizing capability shifts. Six months ago, you had to break tasks into bite-sized pieces because models couldn't hold context or reason through multi-step problems. Now they can. The scaffolding you built to compensate for early limitations has become the ceiling.

> "For modern models, that's actually really not the way to do it."

Cherny's recommendation is simple: describe the task, define success criteria, set boundaries. Then step back. The model needs room to reason, to try approaches you wouldn't have prescribed, to course-correct when something doesn't work. Micromanagement doesn't just waste your time — it constrains the model's ability to find better paths.

Andrew Ng called this "lazy prompting" last year, though the term undersells what's happening. It's not laziness. It's delegation to something finally capable of being delegated to. You add detail only when the model fails without it, not preemptively because you assume it will.

**Key shifts in how to work with capable models:**

- Old way: "Click the submit button, then wait for the response, then parse the JSON"
- New way: "Submit this form and handle any errors appropriately"
- Old way: "First check if the file exists, then read it line by line, then..."
- New way: "Process this file and return the summary"

The gap between human and AI capability hasn't closed — it's inverted in some domains. Coding agents can now hold more context than most developers, reason through edge cases humans miss, and maintain consistency across thousands of lines. The bottleneck isn't the model. It's the manager who can't stop telling it exactly how to do its job.

Cherny works at [Anthropic](https://wire.fourthweb.ai/tag/anthropic/), so he sees model improvements before the rest of us. But his advice carries weight because even [Anthropic's own employees keep discovering new capabilities](https://www.businessinsider.com/anthropic-claude-code-prompting-tips-boris-cherny-micromanaging-ai-2026-7?ref=wire.fourthweb.ai) as they experiment. If the people building these systems are still learning what they can do, the rest of us are definitely underestimating them.

### The Implication

Audit your prompts. Any time you're giving step-by-step instructions, ask if you're solving for a limitation that no longer exists. Test outcome-based prompts against your detailed ones. You'll likely find the "lazy" version is faster and often better.

This matters beyond productivity tips. As models get more capable, the people who figure out how to delegate effectively will build things the micromanagers can't. The shift from "tell it what to do" to "tell it what you need" is the shift from using a tool to managing an agent. The sooner you make it, the more runway you have.

### Sources

[Business Insider Tech](https://www.businessinsider.com/anthropic-claude-code-prompting-tips-boris-cherny-micromanaging-ai-2026-7?ref=wire.fourthweb.ai)