Prompt Engineering for Non-Technical Founders

Why Most Prompts Fail

Three reasons prompts produce mediocre outputs:

1. Too short. “Write a blog post about SEO” is a template of a request, not a request. The model has no context, format, audience, or angle. Result: generic.
2. No examples. Humans learn from examples. So do models. Prompts without examples force the model to guess what “good” looks like.
3. Ambiguous asks. “Make it better” is undefined. Better how? Longer? Sharper? Simpler? Without specification, the model guesses.

The 7 patterns below fix all three.

The 7 Patterns That Work

Pattern 1: Role Assignment

What it is: Start the prompt by telling the AI who to be.

Why it works: Models respond dramatically differently based on assigned role. “You are a copywriter” produces different output than “You are a strategic consultant” for the same downstream task.

Template:

You are a [specific role] with [specific experience]. Your job is to [specific outcome].

Example:

You are a B2B copywriter with 10 years of experience writing landing pages for SaaS companies. Your job is to write a landing page hero section that maximizes trial signups.

Pattern 2: Context Stacking

What it is: Give the AI everything it needs to do the job well; audience, brand, current situation, prior attempts.

Why it works: Models can’t infer context they don’t have. The more relevant context you provide, the more useful the output.

Template:

CONTEXT:- Business: [what you do]- Audience: [who buys]- Current situation: [where you are now]- Prior attempts: [what you've tried]- Voice reference: [3-5 sentences in the voice you want]

Rule: Spend 3x more effort on context than on the actual ask. Context is 80% of prompt quality.

Pattern 3: Explicit Format Specification

What it is: Tell the AI exactly what output format you want.

Why it works: Without format guidance, the AI defaults to prose paragraphs. Sometimes you need a table, a JSON object, a bulleted list, or a specific structure.

Template:

Output format:- Section 1: [what goes here]- Section 2: [what goes here]- Total length: [word count or bullet count]- Style: [tone/voice details]

Example:

Output format:- 5 bullet points- Each bullet: one specific tactic + expected outcome- Total length: under 300 words- Style: direct, no marketing fluff

Pattern 4: Example-Driven Prompting

What it is: Show the AI 1-3 examples of exactly what you want.

Why it works: Examples anchor quality. Models pattern-match on what you show them and produce output that matches.

Template:

Here are 2 examples of the output style I want:Example 1: [paste]Example 2: [paste]Now produce a similar output for [my actual request].

When to use it: Any time you have prior good outputs. Any time you want brand voice matching. Any time you want a specific structure.

Rule: One good example is worth 500 words of instruction.

Pattern 5: Step-by-Step Decomposition

What it is: Ask the AI to think through the problem step by step before producing output.

Why it works: Models produce better final answers when they reason first. Just adding “think step by step before answering” measurably improves output quality on complex tasks.

Template:

Before writing your final response, think through:1. Who is the audience and what do they need to know?2. What's the strongest angle for this argument?3. What examples would make this concrete?4. What could go wrong or be misinterpreted?Then produce your response.

Pattern 6: Explicit Constraints

What it is: Tell the AI what NOT to do, not just what to do.

Why it works: Models default to certain patterns (using jargon, hedging language, generic openings). Explicit “don’t do X” constraints work.

Template:

Constraints:- Do not start with "In today's..."- Do not use "leverage," "delve," "crucial," "it's worth noting"- Do not hedge with "may," "might," or "could be"- Every claim needs a specific example- Maximum 3 bullet points per section

Impact: Adding negative constraints alone often improves output quality more than positive instructions.

Pattern 7: Iterative Refinement

What it is: Treat the first output as a draft. Refine through follow-up messages.

Why it works: The best output is rarely the first output. Iteration produces dramatically better results than trying to perfect a single prompt.

Refinement patterns: - “This is too generic. Add 3 specific examples.” - “Rewrite paragraph 2 to be more direct. Cut 30 words.” - “The opening reads like AI. Rewrite it to sound like [voice reference].” - “Add a counterargument section.”

Rule: 3-4 rounds of iteration typically produce publication-quality output. Trying to nail it in one prompt usually wastes more time than iterating.

Copy-Paste Templates

Template A: Blog post draft

You are a [industry] expert with [X] years of experience writing content for [specific audience].CONTEXT:- Business: [description]- Target reader: [specific profile]- Reader's problem: [what they're trying to solve]- Voice reference: [3-5 sentences from prior content]TASK:Write a [word count] blog post arguing [specific stance] on [topic].STRUCTURE:- 40-60 word direct answer in the first paragraph- H2 sections: [list them]- Include these examples: [list them]CONSTRAINTS:- No generic openings- No "leverage," "delve," "in today's..."- Every claim needs a specific example- Direct, opinionated toneThink through the audience, angle, examples, and risks before writing.

Template B: Email drafting

You are a founder writing a follow-up email to a warm prospect.CONTEXT:- Prospect: [name, company, role]- Prior conversation: [what was discussed]- Voice reference: [3 emails you've sent that landed well]TASK:Draft a 4-sentence follow-up email that:- References something specific from the prior conversation- Offers one specific piece of value- Ends with a soft askCONSTRAINTS:- Under 90 words- No formal openers ("I hope this finds you well")- Sound like a person, not marketing

Template C: Research synthesis

You are a strategic analyst.CONTEXT:- Question: [specific question]- What I already know: [summary]- What I need to decide: [decision context]TASK:Research [specific topic] and produce:1. 3-5 key facts with sources2. Common misconceptions3. What most people miss4. What this means for my decisionCONSTRAINTS:- Every fact needs a verifiable source- Flag statistics you're uncertain about- Distinguish "widely reported" from "actually confirmed"

Debugging a Bad Output

When output is bad, ask yourself:

Is the role missing? Add a specific role.

Is context missing? Add background, audience, and prior work.

Is the format vague? Specify structure, length, and style.

Are examples missing? Add 1-3 examples of what “good” looks like.

Are constraints missing? Add “do not” rules.

Did I ask for one-shot output on a complex task? Break it into steps or iterate.

Most bad outputs are diagnosed by finding which of these 6 is missing.

The 3-Second Test

Before sending a prompt, read it in 3 seconds and ask:

Would a smart intern know exactly what to do?

If yes, the prompt is probably good. If they’d have questions (“what audience?” “what format?” “what tone?”), those questions are your missing context.

Advanced: Chained and Multi-Turn Prompts

For complex outputs, chain prompts across multiple turns:

Turn 1: Ask for an outline.

Turn 2: Ask for section 1 based on the outline.

Turn 3: Ask for section 2, referencing section 1.

Turn 4: Ask for the intro, referencing all sections.

Turn 5: Ask for a final polish pass.

This produces higher quality than trying to generate everything in one shot. The model has less to hold in memory at each step.

FAQ

How long should I spend on a prompt? For a one-off task: 1-3 minutes. For a task you’ll repeat: 30-60 minutes to build a strong template you’ll reuse. The template pays back on the second use.

Do the same prompts work for Claude and ChatGPT? Mostly yes. Both models respond to role assignment, context, examples, and constraints. Small tuning helps, Claude prefers slightly less structured prompts; ChatGPT loves explicit format specification.

How many examples should I include? 1-3. More than 3 rarely helps and consumes context window. Focus on quality examples, not quantity.

Should I use markdown, XML, or plain text in prompts? Any of them works. Consistency matters more than format. Pick a style and stick with it across prompts so you get consistent outputs.

Are there prompt libraries I should buy? Rarely worth it. Most sold prompt libraries repackage the 7 patterns above with industry-specific examples. Building your own template library from your actual use cases produces better ROI.

Key Takeaways

7 patterns handle 90% of prompt engineering needs: role, context, format, examples, decomposition, constraints, iteration.

Context is 80% of prompt quality. Spend most of your prompt time there.

One good example beats 500 words of instruction.

Iteration produces better output than trying to nail a single prompt.

The 3-second test: would a smart intern know exactly what to do?

If you’d like Octo Partners to build a prompt library for your business, including reusable templates for content, sales, and analysis, book a free Strategy Call. We build prompt systems that produce senior-quality output at intern-level effort.