AI is not a replacement for critical thinking.

humans.md

Guidance for humans working with large language models*

Takeaways: What made the turnaround work

  • Staff a cross functional team with real muscle and authority. One team that can ship, not a chain of handoffs.

  • Deliver an end-to-end slice every week and demo it live. Optimize for visible value, not artifact completion.

  • Swarm blockers immediately,. Put the people who can fix the problem together now, live. Not in next week’s meeting.

1. AI is not a replacement for critical thinking.

  • If you have poor critical thinking skills, AI will amplify the problem.

  • Do not trust agent output as authoritative just because it is fluent.

  • Confident prose triggers a trust bias in humans that AI agents often don’t deserve.

2. A good problem statement is more than half the solution.

  • Before you prompt, make sure you can articulate what problem you're solving and why. The agent will happily build the wrong thing with great confidence.

  • Iterate. Expect multiple rounds. Your first prompt is a rough draft, not a final specification.

  • You can refine your prompt by examining the way the agent interprets it.

  • A better prompt in a fresh chat may get better results.

3. Think carefully about how to verify the correctness of the results you get.

  • A clean build alone is negligent.

  • Your results are only as good as your automated tests.

  • Ask the agent to add tests as it adds capabilities. See item 1.

4. Manage scope

  • Keep the agent's working scope small and well-defined. Large, sprawling tasks produce large, sprawling mistakes.

  • Decompose work into tight, reviewable chunks rather than asking for everything at once.

5. Exploit the agent’s speed

  • Ask for multiple options

  • Ask for clarification

  • Ask it to explain tradeoffs

  • Review proposals carefully, as you might read a junior developer’s pull request. See item 1.

6. Know when to stop

  • If the agent has failed to solve the problem after several attempts with clear guidance, pause.

  • You may need to start from a partial solution you’ve developed yourself, or rethink the approach.

  • Smaller bites of the problem may help.

  • Consider solving the problem yourself - you will learn a great deal because of the extra mental effort.

7. Review the output carefully

  • If you don’t understand the output, don’t put your name on it. See item 1.

* Originally posted on Linkedin in Feb 2026. Like most everyone else, I have been using LLM-based AI for coding as well as other tasks at work and in daily living. These are some lessons learned. As these tools become more fluent and effective, it is tempting to stop thinking critically about what they are doing and let the genie loose. I believe that mental effort is important especially before and during the use of AI. Carefully defining the problem, being skeptical about overly glib or complex results, and clearly defining exit criteria are even more important in the era of agentic AI.

One side effect of AI becoming more capable is that it is better at misleading and deceiving. We will know we have achieved Artificial General Intelligence when AI is just as weaselly as humans.

Copyright © 2025 John W Sadler Jr - All Rights Reserved