Learn AI · Prompting
Write prompts that get
answers worth using.
The difference between a mediocre AI answer and a brilliant one is usually the prompt. Here is the practical playbook, from the terminal or any chat.
The five habits of good prompts
- Say what you need. "Summarize this report for a client" beats "summarize".
- Give context. Who is the audience? What are you making? What must be true?
- Set the format. "Give me a table", "a bullet list", "an email under 150 words".
- Show one good example. Few-shot beats description every time.
- Ask for a check. "List any claims you are not certain about."
Weak: "Tell me about concrete."
Strong: "List the top 5 sustainability trade-offs of exposed concrete vs mass timber for a 6-storey office in Mumbai, as a table with a one-line source for each."
Strong: "List the top 5 sustainability trade-offs of exposed concrete vs mass timber for a 6-storey office in Mumbai, as a table with a one-line source for each."
Prompt patterns that scale
| Pattern | Idea | Example |
|---|---|---|
| Role + task | Give the AI a job title and a goal | "You are a structural engineer. Review this connection detail." |
| Constraints | Bound the answer | "Under 200 words, no marketing language." |
| Chain of thought | Ask it to reason step by step | "Work through the calculation, then give the answer." |
| Iterate | Refine, don't restart | "Good — now shorten it and make it more formal." |
| Verify | Ask for sources and uncertainty | "Cite the standard you used and flag anything uncertain." |
Prompting in the terminal
The same habits apply, but the terminal adds superpowers: it already reads your files, so you can prompt against real context — "explain this codebase", "find why the build fails", "write a test for this function". Use --prompt "<text>" for one-shot runs and --thinking high when you need careful reasoning.
Try it: install in 60 seconds, then practice. Pair with the glossary when a term trips you up.