principles.fyi Part 2 / 5

LLMs in the Wild · Part 2 / 5

Talking to it — prompting

You do not retrain a model to make it do a new task. You just ask — and how you ask changes everything.

In Book 01 we made a model better by changing its weights — pretraining, then fine-tuning, then preference-tuning. That is slow, expensive, and permanent.

There is a second knob. A far cheaper one.

Just ask

A pretrained model can pick up a brand-new task with zero weight changes. You don’t open the hood. You just put the right words in front of it.

The text you hand it is a . Finding words that work is . The cheapest version is : no examples, just the question.

Sometimes that’s enough. Sometimes it rambles, or answers in the wrong shape. So you give it a couple of worked examples first — — and it copies them. No gradients, no updates: it reads the pattern straight out of the context. That trick is .

One review to classify, one frozen model. Add worked demonstrations, or state the rule with a system prompt — the answer snaps into shape, and the weights never move. The amber block is the part of the context doing the work right now.

Try it: start with both toggles off — watch it ramble. Now flip Add demonstrations. Same model, but two examples taught it the shape of the answer. The weights are byte-for-byte identical.

The third layer: a role

There’s one more piece of context, and you never see it. Before your message, the app silently glues on a — a few lines that set the model’s role, tone, and rules (“You are a helpful assistant. Reply with one word.”).

Real assistants ship long system prompts: thousands of words of standing instructions, every single turn. Same in-context trick, just always-on.

Two knobs, not one

So why does any of this work without training? Because of Book 01’s fine-tuning stage, — that’s what taught the model to treat the context as a command instead of just continuing the text.

The line to hold onto:

Go deeper: when prompting isn't enough

In-context learning is powerful but bounded. Examples eat your context window, the model can drift on long inputs, and it can’t truly memorize new facts this way. When you need a permanent new skill or private knowledge baked in, you reach back for the weight-changing tools from Book 01 — full , or the lightweight adapters. Rule of thumb: prompt first, fine-tune only when you must.

Same network. Two ways to steer it. Next: where all that knowledge came from in the first place — the data.

Sources · 5
  1. Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., et al. (2020). Language Models are Few-Shot Learners. Advances in Neural Information Processing Systems 33 (NeurIPS 2020). arXiv:2005.14165.
  2. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C. L., et al. (2022). Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems 35 (NeurIPS 2022). arXiv:2203.02155.
  3. Anthropic. System Prompts. Claude Platform Documentation (Release Notes). Accessed 2026-06-25.
  4. Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., & Chen, W. (2021). LoRA: Low-Rank Adaptation of Large Language Models. arXiv:2106.09685.
  5. Min, S., Lyu, X., Holtzman, A., Artetxe, M., Lewis, M., Hajishirzi, H., & Zettlemoyer, L. (2022). Rethinking the Role of Demonstrations: What Makes In-Context Learning Work? In Proceedings of EMNLP 2022 (pp. 11048–11064). arXiv:2202.12837.
Full bibliography →