Prompt design fundamentals

A vague prompt makes the model guess. Prompt design helps you give the model a clear task, the context it needs, the format you want, and the boundaries it should respect.

"Write me a marketing strategy" is a prompt. It is also missing almost everything that would make the answer useful. Which product? Which market? Which buyer? What kind of strategy: positioning, demand, content, pricing, or retention?

The model will still answer. That does not mean it understood the job. It filled in the blanks for you.

Prompt design is the practice of closing that gap. Treat it like writing a good brief for a smart collaborator.

Build every prompt from four parts

Every useful prompt has four parts: task, context, format, and boundaries. You do not need all four in every casual question, but you should check all four before you blame the model.

Here is the same weak prompt rewritten with all four parts:

You are helping a B2B SaaS company plan its next 90 days of content.

Task: Draft a marketing strategy for a product that helps RevOps teams clean CRM data before quarterly planning.

Context: The company sells to VP-level RevOps leaders at mid-market SaaS companies. The product has 30 customers, a small sales team, and strong proof from 5 case studies. The goal is to create pipeline from buyers already searching for CRM hygiene and forecasting problems.

Format: Return a short strategy memo with sections for audience, positioning, content themes, distribution, and risks.

Boundaries: Keep it under 700 words. Do not suggest paid ads, events, or a full rebrand. Every recommendation should connect to the buyer problem.

You still leave room for judgment. You remove the useless guessing.

Choose loose or tight before you write

Prompts serve different jobs. Pick the job before you write the instruction.

Use loose prompts for exploration

Exploration prompts help you find ideas, questions, angles, and options. You want range. You expect to throw away part of the output.

What are five unconventional ways a B2B SaaS company at 4M ARR could grow expansion revenue without hiring more salespeople?

You give the model room to roam. You might keep one idea and discard the rest. That is the point.

Use tight prompts for production

Production prompts support repeatable work. You want consistency across changing inputs: summaries, extractions, report sections, briefs, outlines, and QA checks.

You are a B2B marketing analyst.

Given the company profile and competitor notes below, produce a positioning summary in this exact format:

- Market category:
- Competitive alternative:
- Target buyer:
- Key differentiator:
- Value proposition:

Write one sentence per field. No preamble. Use only the information provided.

Production prompts need structure, examples, and clear boundaries. A middle-ground prompt usually fails both jobs: too narrow for exploration and too vague for production.

Use techniques that reduce guessing

These techniques work because they add information. Skip tricks and magic phrases. The model responds best to clear instructions, good context, and examples.

Ask for the reasoning path you need

Complex tasks get better when the model evaluates the problem in steps you can inspect.

Do not ask for hidden reasoning. Ask for the visible criteria that matter to your decision.

Evaluate these three pricing models for a new analytics product.

For each model, assess:

1. Revenue predictability
2. Ease of implementation
3. Buyer expectations
4. Retention risk

Then recommend the strongest option and give a short rationale.

The answer will take more space. It will also show you where the judgment holds and where you need to challenge it.

Give examples

Examples control output quality better than abstract description. One example teaches format. Three examples teach pattern, tone, and judgment.

Use examples when you care about style or structure:

• Paste two strong research summaries, then ask for the next one.
• Paste three customer interview notes in the format you like, then ask the model to format a new interview.
• Paste a good outreach email and a bad one, then ask the model to explain the difference before drafting.

The model does not need a long lecture about the pattern. It can infer the pattern from examples.

Assign a useful role

Role assignment nudges the model toward a domain, expertise level, and tone.

You are a skeptical CFO reviewing this business case.

That prompt will produce different feedback than this one:

You are a supportive team lead helping improve this business case.

Use roles lightly. A role can shift the lens, but it cannot give the model knowledge it does not have. If accuracy matters, provide the source material.

Specify structured output

Production prompts need predictable shape. Ask for the exact structure you need.

Weak:

Make a table of the risks.

Strong:

Return a markdown table with these columns:

- Risk
- Why it matters
- Evidence from the source
- Severity: High, Medium, or Low
- Recommended action

For software or data workflows, use explicit JSON fields:

{
  "summary": "string, max 100 words",
  "key_risks": ["array of strings"],
  "recommendation": "proceed | revise | abandon",
  "confidence": "number from 0.0 to 1.0"
}

Stack the pieces for production prompts

A strong production prompt often combines all four techniques: role, task, examples, structured output, and boundaries.

Use this template when you need a prompt you can run more than once:

Role:
You are a [specific role] helping [specific audience] make [specific decision].

Task:
Analyze the input below and produce [specific output].

Context:
[Add audience, product, market, source material, constraints, definitions, and anything the model should know before answering.]

Process:
First, identify [criterion 1].
Then, evaluate [criterion 2].
Then, flag [risks, gaps, or assumptions].
Finally, recommend [decision or next step] with a short rationale.

Format:
Return the answer as [markdown memo, table, JSON, bullet list, numbered plan].

Examples:
[Paste 1 to 3 examples if format, tone, or judgment matters.]

Boundaries:
Keep it under [length].
Use only [allowed sources].
Avoid [topics, tone, claims, or approaches].
Call out uncertainty instead of inventing details.

Input:
[Paste the material.]

Each layer removes ambiguity. That makes the answer more predictable and easier to evaluate.

Fix prompts by diagnosing the failure

Expect to revise your prompt. Write it, test it, read the output, diagnose the failure, and change one variable.

Task failure

The model misunderstood the job.

Fix it by naming the action more clearly. Replace "analyze this" with "identify the three strongest objections a buyer would raise and recommend how to answer each one."

Context failure

The model lacked the information it needed.

Fix it by adding source material, definitions, buyer context, constraints, or examples. If the model has to guess the company, audience, or goal, your prompt needs more context.

Format failure

The model found useful content, but used the wrong shape.

Fix it by specifying the output structure. Add a table schema, section headings, JSON fields, or an example.

Boundary failure

The model went too broad, too long, too generic, or into topics you wanted to avoid.

Fix it by adding limits. Set length, source rules, exclusions, and quality bars.

Change one thing at a time. If you rewrite the task, context, format, and boundaries in one pass, you will not know which fix worked.

Avoid common traps

Blaming the model for vague prompts

"The model gave me garbage" often means "my prompt gave the model nothing to work with." Before you blame the answer, audit the input. Did you define the task, context, format, and boundaries?

Over-specifying the prompt

Production prompts need constraints. Too many constraints can remove the part of the answer you wanted: judgment. Specify what matters and leave room where you can.

Copy-pasting prompts across contexts

A prompt that works for one task can fail on a similar task. The audience changes. The source material changes. The goal changes. Reuse the structure, then adapt the specifics.

Confusing prompt design with context

A clear prompt can still fail if the model lacks the right information. "Summarize this document" will fail if the document is not in the conversation or connected source. Prompt design handles the instruction. Context gives the instruction something to work on.

Treating prompts like magic words

Prompt hacks work only when they add clarity or useful context. Skip secret phrases. Write clear instructions.

Practice prompt design

Prompt design gets easier when you stop looking for perfect wording and start looking for missing information.

Further reading

• Anthropic's Prompt Engineering Guide covers prompt structure, examples, XML tags, and structured output.
• OpenAI's Prompt Engineering Guide gives a complementary view from the GPT side.
• Prompt Engineering Guide by DAIR.AI goes deeper on zero-shot, few-shot, ReAct, and other techniques.
• Co-Intelligence by Ethan Mollick teaches the working mindset behind better human-model collaboration.

For longer public curricula, start with Learn. For quick tasks, stay in Guides.