Manuals / Prompt Engineering / Ch 1

StartBeginner25 min read

1. How LLMs actually work

Prompt Engineering · 48 pages source format

Models predict the next token — they do not "know" facts. Treat every output as a draft until verified. This mental model saves you from trusting confident nonsense.

What you'll learn

  • Token prediction
  • Hallucination risk
  • Temperature & sampling

Predict vs retrieve

LLMs generate plausible text. They are not search engines. Ask the same factual question twice — note when answers drift.

Do this now

Ask a model for a citation to a real paper. Verify the URL exists.

Pro tip. If accuracy matters, require sources you can click.

Clear?

Temperature intuition

Low temperature = more deterministic. High = more creative. Match setting to task.

Do this now

Run the same creative prompt at temp 0.2 vs 1.0. Compare.

Clear?

Checklist