Stage 08 — Prompting

The cheapest way to change an LLM’s behavior. No training, no data, no GPU — just better instructions. Prompting alone gets you 80% of the way for many production problems.

Prerequisites

• Stage 07 (modern LLMs)
• Working API access (Anthropic, OpenAI, or local model)

Learning ladder

1. Prompt fundamentals — system, user, assistant; chat templates; context windows
2. Few-shot & chain-of-thought
3. Structured outputs — JSON mode, tool calling, schemas
4. Advanced techniques — self-consistency, ReAct, tree of thoughts, reflexion
5. Sampling & decoding — temperature, top-p/k, beam, speculative

MVU

You can:

• Reach for prompting first; know when to escalate to RAG, fine-tuning, or agents
• Write a robust system prompt that survives adversarial inputs
• Force valid JSON output 100/100 times for a non-trivial schema
• Pick sampling parameters for a given task (creative vs deterministic)

Exercise

Write a prompt that classifies emails into 8 categories with 95%+ accuracy on a held-out set of 100 examples. Then break it with adversarial cases. Iterate.

Hands-on companions

Watch it interactively:

• Sampling Knobs — same real GPT-2 logits, three sliders (T, top-p, top-k), four very different outputs.
• Beam Search Lab — greedy / beam / sample on real GPT-2 distributions.
• Structured Outputs — editable response, live JSON-schema validator, five named “break it” experiments.
• CoT Lab — reasoning-depth slider showing how partial reasoning lands on different (sometimes wrong) answers.
• Few-Shot Lab — how k=0..3 examples change format adherence and label coverage.

Ship the stack:

• /ship/04 — build the eval harness — measure prompt changes against a golden set, replace vibes with numbers.
• /ship/13 — evals in production — A/B prompt testing with statistical rigor (paired t-test, effect-size threshold).

See also

• Stage 09 — RAG — when prompting alone isn’t enough
• Stage 11 — Agents — prompts in a loop
• Stage 13 — Evaluation