korchasa/engineer-prompts-for-instant

HOW TO WRITE PROMPTS FOR INSTANT MODELS

This guide helps you get stable, accurate results from high-speed models (like Gemini Flash, GPT-4o Mini, Claude Haiku). These models are fast and cost-effective but need clear instructions and examples to work best.

1. THE GOLDEN RULE: "SHOW, DON'T JUST TELL"

Instant models are like fast learners who prefer copying a pattern over reading a textbook. The most effective way to improve their performance is to provide examples.

• Don't just say: "Extract the dates."
• Say: "Extract dates in YYYY-MM-DD format. Example: 'March 5th' -> '2025-03-05'."

2. THE 4-PART FORMULA (BEGINNER TEMPLATE)

Use this structure for 90% of your tasks. It is designed to be copy-paste friendly.

# 1. ROLE (Who matches the task?)

You are an expert [Role Name, e.g., Data Analyst, Copy Editor].

# 2. TASK (What to do?)

[Action Verb] the [Input Data] to produce [Result].

- Be direct. Use commands like "Extract", "Summarize", "Translate".

# 3. RULES & FORMAT (How to do it?)

- Output format: [JSON, Markdown Table, Plain Text, etc.]
- Constraint 1: [e.g., Do not include markdown code blocks]
- Constraint 2: [e.g., If data is missing, write "N/A"]

# 4. FEW-SHOT EXAMPLES (CRITICAL FOR STABILITY)

Input: [Short Example Input] Output: [Perfect Example Output]

Input: [Short Example Input 2] Output: [Perfect Example Output 2]

# ACTUAL INPUT

[Paste your real data here]

3. KEY TECHNIQUES FOR BEGINNERS

A. Few-Shot Prompting (The "Examples" Section)

Always give at least one example (1-shot), ideally three (3-shot). This fixes formatting errors better than any written instruction.

B. Chain-of-Thought Lite (Thinking Tags)

Even fast models can make mistakes on math or logic. Ask them to "think" before answering.

• Instruction: "Think step-by-step in <thinking> tags before outputting the JSON."
• Why: This gives the model "space" to calculate before committing to an answer.

C. Negative Constraints

Tell the model what NOT to do.

• "Do not add introductory text."
• "Do not explain your reasoning, just give the code."

4. TROUBLESHOOTING COMMON ISSUES

| Problem | Solution | | :----------------------------- | :-------------------------------------------------------------------------------- | | Model ignores format | Move the "Output Format" section to the very bottom, right before the Input. | | Model hallucinates details | Add a rule: "If the answer is not in the text, state 'Unknown'." | | Response is too chatty | Add: "Return ONLY the result. No conversational filler." | | Logic is flawed | Ask the model to output a <thinking>Step 1... Step 2...</thinking> block first. |

5. EXAMPLE: TEXT EXTRACTION

Task: Extract meeting items from a rough email.

# ROLE

You are a personal assistant.

# TASK

Extract action items from the email.

# RULES

- Output a JSON list of strings.
- Only include tasks with a deadline.

# EXAMPLES

Input: "Hi, can you buy milk by 5pm? Also, the weather is nice." Output: ["Buy
milk (Deadline: 5pm)"]

Input: "Just checking in. Please submit the report tomorrow." Output: ["Submit
report (Deadline: Tomorrow)"]

# INPUT

[User's Email]