The core mental model

• You’re not “asking”: you’re specifying a task interface. Make inputs/outputs explicit.
• Models pattern-match: show the pattern you want (format + examples) and they’ll imitate it.
• Reliability comes from structure: separate instructions, context, and constraints so the model can’t confuse them.
• Token budget is real: include only the context that changes the decision.

A good prompt has 5 parts

You can treat prompts like a template with stable sections. Use headings or delimiters so the model doesn’t blend your policy, context, and user content.

1. Role: what the assistant should be (reviewer, tutor, debugger).
2. Objective: the success criteria in plain language.
3. Context: the relevant data (docs, constraints, examples).
4. Rules: boundaries—what to avoid, how to handle uncertainty, what to ask.
5. Output format: the exact shape you want (JSON schema, bullet list, table).

ROLE
- You are a careful technical writer.

OBJECTIVE
- Convert the notes into a clean lesson.

CONTEXT
- Notes: """ ... """

RULES
- If information is missing, ask 1 clarifying question.
- Do not invent citations.

OUTPUT
- Return Markdown with: title, summary, sections, checklist.

High-leverage patterns

Use ONLY the reference text below.
If the answer is not in the reference, say: "Not found in reference."

REFERENCE:
""" ... """

Convert each item into JSON.

Example input:
"Fix login bug"
Example output:
{"title":"Fix login bug","type":"bug","priority":"high"}

Now convert:
"Add multimodal lesson"

Before answering, verify:
1) You followed the output schema.
2) You included all required fields.
3) You did not add assumptions.
Then provide the final answer only.

Return JSON:
{
  "needs_tool": boolean,
  "tool_name": "search" | "calculator" | null,
  "tool_args": object | null,
  "final_answer": string
}

Reduce hallucinations (practical)

You can’t “prompt” your way out of uncertainty, but you can make the model admit it earlier and verify claims more often.

• Demand sources: “quote from the provided reference” (or explicitly say “no sources available”).
• Force abstention: “If unknown, say ‘I don’t know’ and ask one question.”
• Split tasks: extraction first, then reasoning. (e.g., “extract facts → then answer”.)
• Use retrieval: for real-world facts, prefer RAG over stuffing everything into the prompt.

Answer using ONLY the "FACTS" list.
If FACTS is insufficient, respond with:
1) "Insufficient facts."
2) The single most important missing fact you need.

FACTS:
- ...