Writing Compact Instructions: The Art of the 'Micro-Prompt'

In the early days of prompt engineering, "More is More" was the rule. Developers wrote long, descriptive essays to coax the best performance out of models like GPT-3. With modern frontier models (Claude 3.5, GPT-4o), the rule has flipped: Density is Strength.

A compact instruction isn't just cheaper; it's mathematically more potent. By removing the linguistic noise, you leave more "Attention" bits (Module 1.3) specifically for the core task.

In this lesson, we master the techniques for shrinking instruction sets without losing control.

---

1. The Strategy of "Nouns over Narratives"

Most human language is built on "Narrative" structures (Subject-Verb-Object-Filler). LLMs excel at "Property" structures (Key: Value).

Narrative Strategy (Weak):

"I would like you to please take a look at the input text and identify if there are any mentions of historical figures. If you find one, list their name and their most famous achievement in a very professional tone."

Micro-Prompt Strategy (Strong):

"Task: Entity Extraction. Scope: Historical Figures. Output: {name, achievement}. Tone: Professional."

Why this works:

You've replaced 40 tokens with 12. The model doesn't have to "parse" your politeness. It immediately recognizes the Directives (Task, Scope, Output, Tone).

---

2. Structural Signposts vs. Sentences

Tokens used to describe the location of data are waste.

• Wasteful: "Please find the data from the user which I have placed below this line of text."
• Efficient: Use delimiters like ---, ###, or [INPUT].

graph TD
    A[Instruction Block] --> B[---]
    B --> C[Data Block]
    C --> D[---]
    D --> E[Instruction Block 2]

By using Standard Delimiters, you save 10-15 tokens per prompt that would otherwise be spent on "Directional Language."

---

3. The "Imperative" Verb Rule

Every instruction should start with an Imperative Verb.

• Weak: "You should try to summarize this."
• Strong: "Summarize."

Implementation: The Verb Dictionary

Task	Bad Verb	Good Verb
Logic	"Think about..."	"Analyze."
Data	"Gather info on..."	"Extract."
Text	"Write a piece on..."	"Draft."
Code	"Can you code...?"	"Implement."

---

4. Implementation: Prompt Component Testing (Python)

When building an enterprise prompt, don't just "guess" the length. Create a script to measure the Token Perimeter of each instruction component.

Python Code: Instruction Benchmarking

import tiktoken

def benchmark_instructions(variants: list):
    enc = tiktoken.get_encoding("cl100k_base")
    for variant in variants:
        tokens = enc.encode(variant)
        print(f"[{len(tokens)} tokens] -> {variant[:50]}...")

# Comparing styles
variants = [
    "I would like you to be very characteristically funny and sarcastic in your response.",
    "Instruction: Sarcastic/Funny personality.",
    "Persona: Sarcastic Comedian."
]

benchmark_instructions(variants)
# Result: 
# 16 tokens -> Narrative
# 7 tokens  -> Structural
# 4 tokens  -> Key-Value

---

5. Avoiding "Instruction Redundancy"

If you have a global system prompt, avoid repeating those rules in the user prompt.

Anti-Pattern:

• Global: "Always output JSON."
• User: "Summarize this in JSON."

The fix: Create a Prompt Middleware that sanitizes user queries to remove redundant instructions before they reach the model.

---

6. Token-Efficient Formatting (Symbols over Words)

Use mathematical and programming symbols to replace logical conjunctions.

• -> instead of "results in"
• ! instead of "not"
• & instead of "and"

Dense Constraint: Length < 50w & No Preamble. (7 tokens) Narrative: "Make sure the length is less than 50 words and don't include any introductory text." (18 tokens)

---

7. Summary and Key Takeaways

1. Key: Value: Treat your prompt like a configuration file, not a letter to a friend.
2. Imperative Verbs: Start with strong, direct commands.
3. Delimiters: Use symbols (---) to separate instructions from data.
4. Prune the 'Identity': Use 1-2 words to define a role, not a paragraph of bio.

In the next lesson, Output Length Control Techniques, we look at چگونه to stop models from writing 500 words when you only wanted 5.

---

Exercise: The Compactor

1. Take your current "System Prompt."
2. Rewrite it using only Keywords and Delimiters.
3. Use symbols like | or -> to replace transitions.
4. Verify the token count.

• Did you achieve a > 50% reduction?
• Challenge: Try to get the prompt under 20 tokens without losing any core requirements.

---

Congratulations on completing Module 4 Lesson 1! You are now speaking in Micro-Prompts.