01 · Prompt Playground

What you’ll learn

A prompt is more than a question. Most production prompts have three parts:

1. System message — sets the persona and rules (“You are a concise analyst”).
2. User message — the actual request.
3. Context — trusted source text the model should answer from.

In the demo, toggle the persona and the grounding context on and off. Notice how adding context makes the answer quote real source text instead of guessing — this is grounding, and it’s the single most effective way to reduce hallucination.

How the mock model works

To keep everything offline, our mockLLM doesn’t predict tokens — it ranks the sentences in the context by how much they overlap with your prompt and returns the best ones, prefixed by the persona. It’s a teaching stand-in, but the lesson is real: the prompt and the context fully determine the output.

import { mockLLM } from '@lib/js';

// No grounding — the model has nothing to quote.
mockLLM('Why does grounding reduce hallucination?');

// With a persona + grounding context.
mockLLM('Why does grounding reduce hallucination?', {
  system: 'concise analyst',
  context: [
    'Grounding means giving the model trusted source text to answer from. ' +
    'When the model quotes retrieved context instead of relying on memory, ' +
    'it is less likely to invent facts.'
  ],
});

from _shared.mock_llm import mock_llm

# No grounding — the model has nothing to quote.
mock_llm("Why does grounding reduce hallucination?")

# With a persona + grounding context.
mock_llm(
    "Why does grounding reduce hallucination?",
    system="concise analyst",
    context=[
        "Grounding means giving the model trusted source text to answer from. "
        "When the model quotes retrieved context instead of relying on memory, "
        "it is less likely to invent facts."
    ],
)

Prompting strategies to try

• Be explicit about format. “Answer in one sentence” or “return a bullet list”.
• Few-shot. Show 1–3 examples of the input→output you want.
• Chain-of-thought. Ask the model to reason step by step before answering.
• Ground it. Provide sources and say “answer only from the context above”.

Rule of thumb: if you can’t get a good answer, the fix is usually more or better context, not a cleverer question.