⚡ Activation

💭 in your head · nothing to type

The agent in your project keeps answering in the wrong format. You've already rewritten the prompt three times, added examples, put a template in the context. It got better, but not perfect.

Somewhere nearby you hear the advice: "Just fine-tune the model."

Before agreeing, consider: what exactly needs to change — the model's knowledge, its manner, or your framing of the task?

👁 Reflection

💭 in your head · nothing to type

These three things are treated differently. The prompt fixes the framing. Knowledge is easiest to supply through context. Manner and format are sometimes genuinely cheaper to bake into the weights.

Recall the course's ladder. Module 04 — phrasing, 05 — context and memory, 08 — splitting work across agents and checks.

Which rung of that ladder are you on right now with your task?

💡 Concept

When NOT to fine-tune

Climb the ladder from the bottom and stop at the first rung that solves the task:

1. Prompt — a clear spec, examples in the request, an answer format (module 04).
2. Context and RAG — search your documents for the needed pieces and place them in the prompt.
3. **Tools and agents** — give the model a calculator, search, a checker; split the task into roles (modules 07–08).
4. Fine-tuning — only if rungs 1–3 hit a ceiling and you have data.

Why this order. The first three rungs change in minutes and roll back with a single edit. Fine-tuning needs a dataset, a run, and a fresh evaluation after every change.

And a separate note on facts. Fresh and changing knowledge is more practical to keep in context: update the document and the answer updates. Weights don't update that way.