The Founder's Guide to Prompt Engineering (Talk More, Type Less)

The Biggest Misconception: You Don't Need the Perfect Prompt

Let me kill this myth right now. You do not need to get it right in one prompt.

The internet is full of "ultimate prompt templates" and "the only prompt you'll ever need." They're all rubbish. Prompt engineering isn't about crafting one magical sentence that produces perfect output. It's about having a conversation.

Think about how you'd brief a really smart colleague. You wouldn't hand them a single sentence and walk away. You'd explain the context. You'd answer their questions. You'd look at their first draft and say "close, but change this." You'd iterate.

That's exactly how you should work with AI.

The iterative approach: 1. Give AI a rough brief with lots of context 2. Review what it produces 3. Tell it what's good, what's wrong, what's missing 4. Repeat until you're happy

This approach works better than spending 30 minutes crafting the "perfect" prompt because AI can course-correct in real time. Your first prompt doesn't need to be precise. It needs to be rich with context. Precision comes through iteration.

Context windows in modern AI tools are massive. You're not going to run out. So stop worrying about being concise and start worrying about being thorough. Tell AI everything. Let it figure out what's relevant.

My Seven Go-To Prompts

After a year of daily AI usage, these are the prompts I reach for most often. They're not templates - they're thinking patterns.

1. "Ask me questions. Don't assume anything."

This is my number one prompt and I use it for everything. Instead of trying to write a perfect brief, I let AI interview me. It asks clarifying questions I hadn't thought about. It surfaces assumptions I didn't know I was making. The conversation that follows gives AI far better context than any brief I could write.

2. "Give me all the different options."

Before committing to an approach, I want to see the full landscape of possibilities. AI is brilliant at generating multiple approaches to a problem, each with different trade-offs. I can then pick the one that fits, or combine elements from several.

3. "Can you do this for me? What are the ways you could do it?"

I always push AI to actually execute, not just advise. Often there are ways it can handle tasks directly - using Playwright for browser automation, AppleScript for system control, API calls for integrations. If you don't ask, you'll never know what's possible.

4. "Give me your recommendation with pros and cons."

When AI presents options, I want its opinion too. Not because I'll always follow it, but because it forces AI to evaluate the trade-offs and take a position. The reasoning behind the recommendation is often more valuable than the recommendation itself.

5. "Give me a plan, then ask me to confirm before executing."

For complex tasks, I want to see the full plan before anything happens. This prevents AI from going down a wrong path for 20 minutes before I notice. Review the plan, correct any misunderstandings, then let it run.

6. "Teach me about [topic] as if I know nothing."

When entering a new domain, this prompt gives you a comprehensive overview without assuming prior knowledge. Follow up with "What are the most interesting things?" and "What blind spots might I have?" to go deeper.

7. "What could I have done better?"

After completing any task with AI, ask this. It will identify inefficiencies in your workflow, suggest better approaches, and help you improve for next time. This is the meta-prompt that makes all your other prompts better over time.

Talk, Don't Type: Why Voice Input Changes Everything

Here's a specific example. I needed to write a blog post about SaaS pricing.

The typed prompt (what most people do): "Write a blog post about SaaS pricing strategy for founders."

The spoken prompt (what I actually did): "I want to write a blog post about SaaS pricing strategy. The audience is early-stage founders who are launching their first product and have no idea what to charge. The biggest mistake I see is founders undercharging because they're scared of losing customers. But the truth is that low prices actually signal low value. I coached a founder last month who was charging $10 a month and couldn't figure out why nobody took the product seriously. When she raised it to $79 a month, conversions actually went up. I want the article to cover value-based pricing versus cost-plus pricing, the psychology of pricing, common mistakes, and how to run pricing experiments. My voice is casual and direct, I lead with personal stories, and I use Australian English."

The second prompt took me 45 seconds to speak. The first took 10 seconds to type. But the output from the second prompt is ten times better because AI knows exactly what I want, who it's for, what angle to take, and what tone to hit.

Wispr Flow is what I use - it sits in the background and converts everything I say to text, right into whatever app I'm using. But there are other options: macOS Dictation (built-in and free), Google voice typing, Otter.ai, even just recording a voice memo and pasting the transcript. The tool doesn't matter. The habit does.

When you speak, you naturally provide more context. You ramble. You go on tangents. You add examples. All of that "noise" is actually signal. AI can parse through a paragraph of rambling speech and extract exactly what it needs far more effectively than it can work with a sparse typed prompt.

The Knowledge Base Advantage

This is the single biggest lever in prompt engineering that almost nobody talks about.

Context is everything. The difference between generic AI output and genuinely useful AI output is almost always the quality of the context you provide.

I've built what I call the Batko Brain - a database of everything I've ever written or said publicly. 800K+ words across LinkedIn posts, Substack articles, podcast transcripts, meeting notes, personal reflections. When I ask AI to write something in my voice, it doesn't guess. It references thousands of examples of how I actually write.

You don't need something that elaborate. But you do need a knowledge base. Here's the minimum viable version:

Option 1: A CLAUDE.md file

If you're using Claude Code, create a CLAUDE.md file in your project directory. Put in your writing style notes, your preferences, your common instructions, your context about the project. Claude Code reads this file automatically and applies it to every interaction.

Option 2: A GitHub repository of context

Create a private repo with your key documents - voice guides, frameworks, common instructions, past examples of good work. Reference it when starting new conversations.

Option 3: Pin your best examples

At minimum, paste 3-5 examples of output you loved into your prompt. "Here are examples of the quality and style I'm looking for. Match this." This alone will dramatically improve output quality.

The principle: AI with your context beats AI without your context, every time. The investment in building a knowledge base pays for itself within the first week. Every interaction after that benefits from the context you've built.

Advanced Patterns: Making AI Do More

Once you've mastered the basics, these patterns take your AI usage to another level.

The options-then-execute pattern

Instead of asking AI to do one thing, ask it to present options first, then execute your choice. "Give me three different approaches to building this feature, with pros and cons of each, and your recommendation." Review the options, pick one (or a hybrid), then say "go with option 2 but incorporate the error handling from option 3."

The reflection pattern

After AI completes a task, ask: "What could I have done better in how I asked for this?" and "How could you use fewer tokens to achieve the same result?" and "Look at the last week of our conversation - what should I have done differently?" These meta-prompts create a feedback loop that makes you a better prompter over time.

The constraint pattern

AI often defaults to the most complex solution. Add constraints: "Do this using only free tools." "Build this without any paid APIs." "Find a solution that doesn't require any external dependencies." Constraints force creativity and often produce simpler, more elegant solutions.

The delegation pattern

Don't just ask AI to advise - ask it to execute. "Don't just tell me what to do, actually do it. Including the manual steps." Tools like Claude Code can use Playwright for browser automation, write and run scripts, interact with APIs, and handle tasks that most people assume require manual intervention. You'll be amazed at what it can do if you simply ask.

The pre-processing pattern

This is a tip from the AI Roadshow I ran across five cities. Pre-process your prompts, especially the first words, as they often follow predictable patterns like "Can you please..." Structuring your prompt with the task first and context second helps AI respond faster and more accurately. Lead with what you want, then add the why and the context.

The Prompt Engineering Improvement Framework

Getting good at prompting isn't about memorising templates. It's about building a habit of reflection and iteration.

Daily: Ask "what could I have done better?"

At the end of each day, look at your AI interactions. Which ones produced great output? Which ones required multiple rounds of correction? What was different about the prompts? Over time, you'll notice patterns in what works and what doesn't.

Weekly: Review your conversations

Every week, spend 15 minutes looking at your best and worst AI interactions. Ask AI: "Look at these conversations. What patterns do you see in the prompts that got good results vs the ones that didn't?" This is genuine self-improvement in real time.

Monthly: Update your knowledge base

Your context evolves. Your preferences change. New tools emerge. Update your CLAUDE.md file, your voice guide, your examples. The knowledge base is a living document, not a set-and-forget.

The progression:

Most people go through three stages of prompt engineering maturity:

Stage 1: Commands - "Write me a blog post about X." Short, generic, and produces generic output. This is where 90% of people stay forever.

Stage 2: Conversations - Rich context, iterative refinement, follow-up questions. Output is significantly better. Most power users reach this stage within a few weeks.

Stage 3: Systems - Knowledge bases, automated workflows, self-improving agents. AI is embedded in your operating system, not just a tool you use occasionally. This is where the real leverage lives, and it's where surprisingly few people have arrived.

The jump from Stage 1 to Stage 2 is the biggest improvement. And the only thing it requires is talking more and typing less. Give AI your full context. Let it ask you questions. Iterate on the output. Don't try to nail it in one prompt.

Quick Reference: Bad Prompts vs Good Prompts

Here are real examples showing the difference between prompts that produce garbage and prompts that produce genuinely useful output.

Bad: "Write me an email to investors." Good: "I need to send an update email to our 12 angel investors. Last month we hit $30K MRR, up from $22K. We hired two engineers. Our main challenge is that churn increased from 3% to 5% and we're investigating why. The tone should be honest and transparent - I always share the bad with the good. Keep it under 300 words. Here's an example of a previous investor update I sent that they responded well to: [paste example]."

Bad: "Make this code better." Good: "This function handles user authentication but it's slow - taking 2-3 seconds on each login. I think the issue is the database query. Can you identify the performance bottleneck and suggest a fix? We're using PostgreSQL and the users table has about 50,000 rows. Here's the current code: [paste code]."

Bad: "Give me marketing ideas." Good: "I run a rostering SaaS for hospitality businesses in Australia. We have 200 customers paying $80/month. Our best acquisition channel is word of mouth but it's too slow. I've tried Google Ads but the CPA was $340 which is too high for our LTV. What are five marketing strategies specifically suited to B2B SaaS selling to small restaurant owners? For each, give me the estimated cost, time to see results, and what makes it work for this specific market."

The pattern is always the same: more context, more specifics, more examples. The extra 60 seconds you spend adding context saves you 10 minutes of iterating on bad output.