Prompt Garage

Findings

• No role defined

  Prompts perform better when they assign the model a clear role.

  Fix: Start with: 'You are a senior <profession> who …'

• Output format not specified

  The model has to guess the output shape — answers will drift.

  Fix: Add: 'Return JSON: { … }' or 'Format: numbered list of …'.

• No constraints

  Without constraints, output length, tone, and content vary wildly.

  Fix: Add: max length, forbidden words, tone, audience.

• No examples provided

  1–2 examples (few-shot) reliably improve output quality.

  Fix: Add: 'Example input: … Expected output: …'

• Prompt is very short

  Only 9 words. Most strong prompts are 60+ words.

  Fix: Add role, context, task, constraints, output format.

• Vague adjectives

  Found vague words: something, good.

  Fix: Replace with concrete, measurable criteria.

• Politeness noise

  Politeness words add tokens without changing output quality.

  Fix: Remove 'please/kindly/thanks'.

• No success criteria

  Telling the model how you'll judge output sharpens results.

  Fix: Add: 'A good answer is one that … '.