Prompting fundamentals

Learn how to write prompts that give an LLM the role, context, format, and constraints it needs to produce useful work.

A good prompt gives an LLM the same things you would give a sharp teammate: a goal, context, examples, constraints, and a clear definition of done. If the answer feels vague, check what the model had to guess.

Use this guide when you want to turn a loose request into a prompt that produces useful work.

Know what an LLM needs

An LLM predicts text from the context you give it and the patterns it learned during training. It can summarize, draft, classify, brainstorm, rewrite, and spot gaps, but it needs clear direction.

Give it these pieces:

Goal: What do you want done?
Role: What perspective should the model take?
Audience: Who will read or use the answer?
Context: What facts, source material, constraints, or examples matter?
Format: What should the output look like?
Quality bar: What would make the answer good enough to use?

Know the surface you are using

Chat products, custom assistants, and raw API calls do not behave the same way. Before you prompt, check what the surface can access.

Surface	What it may include	What to do
Chat products	Browsing, file uploads, memory, connectors, and model choices	Use files and conversation context when the product supports them.
Custom assistants	Approved docs, templates, actions, and saved instructions	Give the assistant the task and the missing context it cannot infer.
API calls	The exact text and data your system sends	Include every fact, example, and constraint the model needs.

If a tool cannot browse the web, read your database, or remember prior work, it may answer without evidence. Give it source material or ask it to say when the answer is not in the provided context.

ChatGPT and Claude give a practical comparison. Use ChatGPT when you need structure, task breakdown, brainstorming across large context, or broad model choice. Use Claude when you need long-document work, writing, ideation, or code. Test each model on the work you care about.

Use a simple prompt shape

State the role. Tell the model what perspective to use, such as "Act as a product marketing lead" or "Act as a skeptical buyer."

Name the task. Use a specific verb: summarize, rewrite, compare, classify, draft, critique, or plan.

Add context. Paste the source material, audience, constraints, facts, and examples the model should use.

Set the format. Ask for bullets, a table, a memo, a checklist, a draft, or another concrete shape.

Define good. Give the model the quality bar, such as length, tone, evidence standard, or what to avoid.

Start with this template:

Act as [role].
Your task: [specific task].
Audience: [who will read or use this].
Context: [facts, source material, constraints, and examples].
Output: [format, length, and sections].
Quality bar: [what good looks like].
If anything material is missing, ask up to [number] clarifying questions before you answer.

You do not need the perfect prompt on the first try. Read the answer, find what the model misunderstood, then tighten the prompt.

Example prompt

Weak prompt:

Write a launch post for our new feature.

Stronger prompt:

Act as a product marketing lead at a B2B SaaS company.

We are launching a feature that flags stalled onboarding accounts before renewal risk grows.

Audience: customer success leaders at mid-market SaaS companies.

Pain point: teams notice onboarding problems too late because account notes, product usage, and support tickets live in separate tools.

Value: the feature shows which accounts need attention and why.

Output: write a LinkedIn post in a clear, non-technical tone. Use short paragraphs. Start with a concrete problem. Include one sentence on the business impact and end with a simple CTA to read the launch note.

Quality bar: avoid hype, avoid jargon, and make the post sound like a person who has seen onboarding problems firsthand.

This works because the model gets a role, audience, pain point, value, format, and quality bar. It no longer has to invent the company, the buyer, or the message.

Make weak prompts stronger

Use this pattern when you want better output: add role, audience, constraints, format, and the reason the work matters.

Marketing

Beginner:

Write three ad variations for new pet owners.

Better:

Act as a copywriter for a premium pet wellness brand. Write three 50-word ad variations for new dog owners on Instagram: one funny, one sentimental, and one direct. End each with a clear CTA.

Strong:

Act as a senior copywriter for a premium pet wellness brand. Write three Instagram ad variations for new dog owners who see their pets as family. Each ad should include a hook, a 2-line body, a CTA, and three hashtags. Give each version a different angle: playful, emotional, and results-focused.

Sales

Beginner:

Roleplay as a skeptical prospect.

Better:

Act as a skeptical marketing operations lead at a mid-market SaaS company. I will introduce our lead scoring product. Challenge me for three rounds on integration work, pricing, and data accuracy.

Strong:

Act as a skeptical buyer at a B2B SaaS company that has wasted money on overhyped tools. You care about HubSpot integration, measurable ROI within 90 days, and clear data logic. I will pitch you. Ask high-friction questions that expose weak claims.

Operations

Beginner:

Create an onboarding checklist from this task list.

Better:

Here is a task list: [paste task list].

Create an onboarding checklist grouped by phase: pre-kickoff, week 1, and week 2. Include internal tasks, client-facing tasks, dependencies, and sign-off points.

Strong:

Act as an operations lead onboarding a new B2B client. Based on this task list, build a 3-phase onboarding checklist: internal prep, client kickoff, and week 1 execution. For each task, include owner, suggested due date, internal or client-facing label, dependency, and blocker to flag. Format as a table.

Product

Beginner:

Suggest edge-case tests for image uploads.

Better:

Generate 10 edge-case tests for an image upload feature with tagging. Include unsupported file types, slow connections, duplicate tags, permissions, and mobile interruptions.

Strong:

Act as a QA lead. Build a test matrix for an image upload and tagging feature. Cover file size, format support, concurrent uploads, tagging errors, permission conflicts, and mobile versus desktop behavior. For each test, include test name, risk level, expected result, and what failure would tell the team.

Use intermediate moves

Once the basic prompt works, add one of these moves.

Ask for a plan first: "Before drafting, outline your approach in 5 bullets."
Give examples: "Use these two examples as the style target. Match structure and level of detail."
Match voice with samples: "Rewrite this using the voice in the samples below. Keep the meaning, but match sentence length, vocabulary, and rhythm."
Transform formats: "Turn this product description into a 5-email sequence for new trial users."
Ask for critique: "Review this draft against the quality bar. List the 5 most important fixes before rewriting."

For voice matching, public writers with a large body of work are easier for a model to approximate. For a company voice, provide voice samples inside the prompt.

Troubleshoot bad output

The answer is bland

Ask the model what it needed but did not have.

Given my prompt and your answer, what context would help you produce a stronger version? List the missing inputs, then write a better reusable prompt.

You can also ask for range:

Give me three versions: safe, bold, and unexpected. Explain when each version would work.

The answer repeats itself

Ask for controlled variation.

Rewrite the same idea in three styles: plain, story-led, and opinionated. For each version, explain the best channel: website, email, or social.

The model misunderstands the task

Make it restate the assignment before it writes.

Before answering, summarize what you think I want in one sentence. Then list your assumptions and ask any clarifying questions that would change the output.

The model invents facts

Constrain it to the source.

Use only the source material pasted below. If the source does not include a fact, write "Not in source." Do not infer numbers, customer names, quotes, or claims.

Know the limits

LLMs can help you move faster, but they do not remove your judgment.

Verify facts, names, numbers, quotes, and sources before you publish.
Keep sensitive data out unless your setup has the right security controls.
Watch for bias, stale assumptions, and confident claims without evidence.
Treat the model as a collaborator, not the decision-maker.

Keep a prompt library

Save prompts that work. A useful prompt library includes:

The job the prompt performs.
The inputs it needs.
The full prompt.
One good example output.
The situations where it works.
The last date someone tested it.

Keep the library small. Use the prompts you trust, improve them when they fail, and delete the ones no one uses.

Prompting fundamentals

On this page