Prompt Engineering Roadmap

Learn to get dependable results from language models: clear instructions, examples, structured output and honest evaluation.

9 milestones · ~86 hours total

1. 1

   How language models behave

   ~6h

   Tokens, context windows, temperature, why models hallucinate, and what they are and are not good at.

   Practice this →
2. 2

   Writing clear instructions

   ~8h

   State the task, audience, format and constraints. Give context, not just a command, and say what to do rather than only what to avoid.

   Practice this →
3. 3

   Examples and structure

   ~8h

   Few-shot examples, role and context framing, delimiters and XML-style tags to separate instructions from data.

   Practice this →
4. 4

   Reasoning and step-by-step tasks

   ~10h

   Ask for working before an answer, break big tasks into a chain of smaller prompts and check intermediate results.

   Practice this →
5. 5

   Structured output

   ~10h

   Request JSON or a fixed schema, validate it in code and handle malformed replies with retries.

   Practice this →
6. 6

   Grounding and reducing hallucination

   ~10h

   Provide source text, require citations or quotes, allow 'I don't know' and verify claims against the source.

   Practice this →
7. 7

   Evaluating prompts

   ~12h

   Build a small test set, define what a good answer is, compare prompt versions and track regressions rather than judging by feel.

   Practice this →
8. 8

   Safety and prompt injection

   ~8h

   Treat user and web text as untrusted, separate instructions from data and limit what a model is allowed to do.

   Practice this →
9. 9

   Prompts in products

   ~14h

   Version prompts, keep them out of scattered strings, log inputs and outputs responsibly and control cost and latency.

   Practice this →