1. Mental Models for Working With LLMs

Why mental models matter

If you treat an LLM like a search engine or a senior engineer on autopilot, you’ll get the worst results: confident-sounding output that drifts, breaks, or hides mistakes. A simple mental model makes your prompting clearer and your verification sharper.

How to use this section

Read these pages with an engineering goal: improve the quality of your next iteration loop.

• Skim first: get the vocabulary (tokens, context window, temperature, tools).
• Then apply: use each concept immediately in a small prompt → code → run cycle.
• Collect patterns: keep a short list of “prompts that work” and “failure modes I hit.”

Skills you’re building

• Calibration: knowing what the model can do reliably vs what must be verified.
• Instruction control: writing constraints that stick and avoiding contradictory context.
• Context discipline: including only what helps, summarizing state, and staying under budget.
• Determinism control: understanding variability and when you want repeatability.
• Verification habits: treating hallucination as normal and building checks into your loop.

Section map (1.1 → 1.5)

• 1.1 What the model is actually doing
• 1.2 The difference between knowledge, reasoning, and tools
• 1.3 Context windows and why “just add more text” fails
• 1.4 Temperature, randomness, and creativity knobs
• 1.5 Hallucinations: the predictable failure mode

A quick practice loop

Use this loop while reading the section. It keeps learning grounded and avoids “prompt theory.”

1. Pick one concept (e.g. temperature, context budgeting).
2. Create a tiny task (e.g. “generate a function + tests,” “summarize this text into JSON”).
3. Run two variants (change one knob or constraint).
4. Verify (tests, lints, runtime output, schema validation).
5. Write one sentence: “When I changed X, Y happened; next time I’ll do Z.”

Where to go next

• Start with 1.1 What the model is actually doing
• Next section: 2. Vibe Coding as a Software Development Style