98 · Letters to a Machine

For questions to ChatGPT, quality matters far more now. It's about the questions you ask and the specific things you want out of one interaction. Stop thinking of it as a fast-paced dialogue — it's more of a slow, tempered mail correspondence you're deriving from LLMs. The questions formulated in your prompts have to be less whimsical and light, more nuanced and detailed. What factors do you want the AI to consider? Give more context. Give more detail. And specify what you actually want it to provide.

Treat LLMs as G-mails with smarter people, not random chats. Providing specific requirements and going deeper gives you edge. Your use of AI should be more like mailing a professor than having a conversation: it needs background, tailored questions, specific requirements. You're not having a random chat — you're leaning into what you want and what you want it to provide. The questions should be lengthy, not an emotional lash-out.

Don't use AI at surface level. Even when it gives great information — which it mostly does — don't just pace through the answers. Break down what each piece is, regurgitate it, process it. The swing should be slow. Give polished, tailored, deeply-thought-out questions, and the answers correspond to that effort.

I'm repeating this because it matters: use AI the way you'd mail someone brilliant. Don't throw snippets of cheap information. Struggle to write one coherent text that specifies exactly what you want done and what the current bottleneck is.

A worked example — the prompt that actually moved my physics: here is the past paper, here is the marking scheme. One: give me the model answer — the most concise, succinct version. Two: tell me the thought process to write this out — the cognitive algorithm of reactions; as you read the question, which terms did you recognise and why did you react to them? Three: which command terms triggered which thought process? That's a prompt with a job description. The generic version gets generic mush.

Another: I have uploaded a past paper and its mark scheme. Do NOT solve the paper. Instead, extract a list of mark-scheme rules — specific phrases, steps and formats required to award marks. For each rule: the exact required phrasing, what students typically write instead that loses marks, and how many marks are lost per omission. Output as a table. Be ruthless — but confined to the A-level marking schemes, and only rules that repeat across questions.

DeepSeek is fast and adaptive — responsive, good general advice instantly, but low throughput for deep thought. For processing large datasets and analysis, Claude is the champion. You don't need many LLMs; don't overcomplicate it. Use two, properly.

The division of labour is the point: AI does what it does best — code, drafts, consulting-shaped synthesis — and you do what you do best: the intensity, the judgment, the vows. I don't think any chatbot really understands the intensity we're operating at. Don't ask it for plans; ask it for contents, select from those, and go deep on concrete conversations — fixing errors, filling knowledge gaps. For conjuring the work ethic, you are too powerful for it.

Just got a DM from Albert — maxed out Claude, here I come. What an instantaneous action: mid-chat, bought it, done. Upload all the data, tell it to go through every revision note on PMT and the past papers, and generate only the key information for maximum output. That's the loop.