How to make a probabilistic model behave like a reliable business tool. Last week you drove a coding agent and saw its harness. This week we zoom into one layer (the instruction layer) because it is the highest-leverage, lowest-cost skill in the entire stack.
Same model. Same 20 messy vendor emails. The only thing that changes is the prompt. (Illustrative teaching example; you’ll reproduce this effect yourself in the hands-on.)
When an LLM feature underperforms, there is an ordered menu of fixes, and the price rises steeply as you go down it.
| Fix | Cost | Speed | When |
|---|---|---|---|
| 1 · Edit the prompt | Free | Seconds | Always start here |
| 2 · Add few-shot examples | Free (a few tokens) | Minutes | Format or edge-case misses |
| 3 · Better context / retrieval | Engineering time | Days | Model lacks the facts (Week 7: RAG) |
| 4 · Fine-tune | Data + compute + upkeep | Weeks | Consistent style at scale, hundreds+ examples |
| 5 · Wait for a new model | — | Quarters | Not a strategy |
Everything the model knows about your task has to fit in one window. Press Play and watch the pieces arrive: system prompt, examples, retrieved documents, chat history, the new question. When the window is full, the oldest turns are evicted, and the answer quietly changes.
Try this: in “Overflow”, note which piece carried the fact the answer needed. Then in “Curated”, say in one sentence what the summary step traded away to keep it.