Diksha Anand

Posted on Aug 26

What are the Basics of Prompt Writing ??

#ai #productivity #beginners #machinelearning

Prompt writing (prompt engineering) is the practice of structuring instructions so an LLM generates precise, high-quality, and contextually accurate outputs. Rather than treating the model like an open-ended conversationalist, effective prompting treats it like a programmable engine that requires clear constraints, context, and schemas.

Component	Purpose	Example
1. Role / Persona	Sets domain expertise, perspective, and technical level.	"Act as a Principal Database Architect specializing in distributed systems."
2. Task / Objective	States the exact deliverable in one direct sentence.	"Compare PostgreSQL and MongoDB for high-concurrency write workloads."
3. Context & Input Data	Feeds raw data, constraints, target audience, or source copy.	"Audience: Staff Engineers. Target workload: 10k writes/sec with JSON payloads."
4. Negative Constraints	Explicitly bans unwanted behaviors, filler words, or patterns.	"Do NOT use generic intro phrases ('In today's world...'). Do not use buzzwords like 'seamless'."
5. Structural Schema	Mandates the output format (Markdown table, JSON, bullet points).	"Output a 4-column Markdown table: Feature, Postgres, MongoDB, Primary Trade-off."
6. Examples (Few-Shot)	Anchors expected tone, length, and formatting style.	Provide a 2-sentence golden example of the expected technical depth.

The Universal Prompt Architecture
Here is the modular framework to structure everyday professional prompts:

# ROLE & OBJECTIVE
Act as a [Specific Role/SME]. Your task is to [Specific Task/Deliverable].

# CONTEXT & INPUTS
- Target Audience: [e.g., Beginners / C-Suite / Technical Specialists]
- Source Data / Background: [Insert raw facts, parameters, or text]

# OPERATIONAL GUIDELINES
1. [Core instruction 1]
2. [Core instruction 2]

# NEGATIVE CONSTRAINTS (What to Avoid)
- Do NOT include conversational throat-clearing (e.g., "Sure, I can help with that", "Here is...").
- Do NOT summarize or repeat the prompt premise.
- Avoid generic adjectives; quantify claims with specific units or bounds.

# OUTPUT FORMAT
[Specify exact layout: JSON schema, Markdown tables, or numbered steps]

Core Prompting Techniques

1. Zero-Shot vs. Few-Shot Prompting

Zero-Shot Prompting:

Asking the model to perform a task using instructions only. Best for straightforward and well-defined tasks.

Few-Shot Prompting:

Including 1–3 high-quality input/output examples inside the prompt. Best for strict formatting, nuanced tone matching, or complex classification.

2. Chain-of-Thought (Reasoning Steps)

Adding instructions such as "Think step-by-step before providing your final answer" can encourage more deliberate reasoning and may improve performance on complex logical or mathematical tasks.

Note: For production prompts, you generally don't need to request hidden reasoning. Instead, ask for the key reasoning, assumptions, or verification steps you actually need to see.

3. Chunking (Modular Prompting)

Instead of asking the model to generate a 3,000-word document in one prompt, break the workflow into sequential passes:

Outline → Section Drafts → Editorial QA

This gives you more control over quality, consistency, and revisions.

Common Prompting Mistakes & Fixes

Mistake 1: Being Too Broad

Bad:

"Write about SEO."

Fix:

"Create a 5-step checklist for diagnosing indexation drops using Google Search Console."

The second prompt defines the topic, task, format, and scope, giving the model much less room to guess.

Mistake 2: Telling the Model What to Do, But Not What to Avoid

Problem:

Without negative constraints, AI may introduce clichés, unnecessary filler, repetitive transitions, or generic language.

Fix:

Add explicit constraints such as:

"Avoid AI clichés, generic introductions, unnecessary repetition, and robotic transitions."

Mistake 3: Forgetting the Output Structure

Problem:

A vague output request can produce inconsistent formatting.