ai · llm · context

Prompting is dead. Long live context engineering.

"Prompt engineering" made it sound like the trick was the perfect sentence. In practice the trick is the perfect pile of stuff — what the model sees, in what order, and how much of it.

A checklist I'd actually use

1. Say the job plainly. Role, goal, constraints, output format. Boring beats poetic.
2. Show, don't just tell. Two good examples outperform a paragraph of rules.
3. Retrieve, don't dump. Put in the three relevant paragraphs, not the whole wiki.
4. Keep tools small and named well. The model chooses tools by their descriptions.
5. Prune history. Summarise old turns; drop stale tool output.
6. Measure. If you cannot test it, you are just vibing at a text box.

Prompts did not die — they just got promoted from a sentence to a system. Treat the window like a budget, write it like an API contract, and test it like code.