How to Think with AI, delivered by Rob Rawson
The views I express here are mine alone and do not necessarily reflect the opinions or official position of my employer.
Right before the Easter holiday break, Rob gave a short presentation on how to think with AI and how to use it well. In just five minutes, he broke the topic down into a few clear and practical ideas.
1. The problem: LLMs are prediction machines
LLMs optimise for plausibility, not accuracy or truthfulness.
That framing was helpful. It explains why AI can sound confident while still being wrong. I recently read several articles, including Quantifying uncert-AI-nty: Testing the accuracy of LLMs’ confidence judgments, that made the same point: AI excels at sounding plausible, but its confidence doesn’t always match reality. Rob’s summary reinforced this, while AI can be useful, but it shouldn’t be treated as a source of truth by default.
2. AI amplifies experts. It does not create them.
The real gap is not prompt technique. It is the domain knowledge behind the prompt.
This really resonated with me. Recently, I was talking to a friend who works as a tech project manager. I showed her how to bring an MVP to life using prompts, and at the end of our session, she was amazed by how advanced AI has become. She asked, "If I can vibe-code any app, what’s the need for a software engineer?"
My response was simple: do you know if you have accidentally committed secret credentials to GitHub? Do you know how to scale an application to 100 or 1,000 users? Do you know how to deploy it safely and not just on your local machine? Do you know how to handle race conditions?
She paused.
That is the knowledge gap. Vibe coders can absolutely learn these things, but the point remains: AI does not replace expertise. It reflects and accelerates the expertise already present. Writing strong specifications, iterating well, understanding trade-offs, and building reliable systems still depend on real experience and judgment.
As engineers, we need to adapt quickly, keep learning, and deepen our understanding of system design and in our domain expertise so we stay ahead of the curve.
3. Offload cognitive load to AI
Your working memory is the bottleneck, so outsource it.
The skill is not just using AI. The skill is knowing what to delegate and what to keep for yourself. Judgment, taste, and trade-offs still matter. Which tasks belong to your own thinking, and which can be offloaded to AI?
This reminded me of how difficult it can be to move from being an individual contributor to becoming more of an orchestrator. That shift was never easy, and AI makes it even more visible. Execution still matters, but delegation becomes part of the craft.
4. Treat output as v0.1, not v1.0
AI output should be treated as a draft, not a finished product.
That mindset is important. The first response is usually good enough to get started, but rarely good enough to ship. The real value comes from engaging with the LLM: asking follow-up questions, interrogating its answers, reviewing and extracting what’s missing, and then refining and improving the output.
5. Close the Understanding loop
Don't just merge the solution.
Complete a "reverse-knowledge transfer" to truly understand and internalize the changes:
- Explain it back: Summarize the changes. Why was this approach chosen?
- Evaluate trade-offs: Compare this solution to alternatives. What are the pros and cons, what are the limitations?
- Generalize: Identify where else in the codebase this principle or pattern could apply.
- Codify: Document the key insights in your CLAUDE.md or team knowledge base.
After any incidents, make sure to implement these steps and commit your lessons learnt. Otherwise, you risk repeating the same mistakes.
Outro: Keep applying what you learn
According to a LinkedIn post I read recently: the half-life of technical skills is now under 5 years, and with the AI boom, it is even shorter. Knowledge from a course often fades if it is not applied to a real project within 90 days.
That feels true.
So over the next 90 days, I want to put these condensed notes into practice rather than just agreeing with them in theory.
One final thought: it is amazing how tools like Claude can be used most effectively when they are treated as collaborators for thinking, not replacements for it. In this day and age of AI, the ones who utilizes it to its maximized potential are going to be the winners in this game.