After many experiments and conversations, it is clear that prompt engineering is not about writing smarter prompts.
Why it matters: Many of us begin with simple requests and are pleasantly surprised by the results.
- However, issues arise when we require something specific, work-related, or of higher quality.
- At that point, obtaining practical results can be challenging.
Between the lines: The problem doesn't lie with the AI model, training data, or the commands we use.
- Instead, it is rooted in the fundamental difference between AIs and real life, specifically in the context—the ability to perceive and understand the surrounding world.
- For instance, a dog knows that humans have two hands, whereas an AI may not grasp that concept.
- Similarly, your coworker understands the appropriate tone for writing a memo to your boss, while the AI lacks that nuance.
Zoom in: This is why simply adding modifiers like "Do not sound like an AI" is not enough.
- Many people think they should instruct an AI as if it were a junior associate or intern.
- While this approach may generate text that sounds more natural, it often results in content that lacks a professional quality.
Best practice: However, the idea of describing in detail what to do and not to do is the first lesson toward obtaining consistent results with AI. The three main rules are:
- Describe the tone, not the output. For example, 'Write like a 50-year-old man who lost everything in the 2008 crash, ranting at a cafeteria over spilled coffee.' This provides context in the form of artificial memories.
- Set the scene as if describing a play to someone who is blind. For instance, instead of asking the AI to 'Write for a website using H3 headings,' prompt, 'You are publishing this on LinkedIn. Make it scannable. Think like a person who just woke up and is reading while sipping coffee.'
- Allow for imperfection. The reason is that AI tends to go overboard when asked for perfection. However, when you ask for messy, raw, or flawed, it becomes much more believable. For example, 'Draft a brain dump after a long day. Do not worry about following every grammar rule. Just get the ideas down.'
Examples: Begin keeping a file of your best prompts for reuse and sharing with friends and coworkers. Here are a few of the most effective ones I've learned:
- "Write this as a 5-year-old."
- "Explain this to a non-technical CEO who keeps asking the same question."
- "You are writing this post for Reddit. Be brutally honest."
- "You are trying to convince your skeptical neighbor to try this app."
- "Draft this as a voice note you're about to send to your boss."
These prompts help eliminate the artificial, overly descriptive, and awkward-sounding phrases that AI often introduces.
Conclusion: AI is not a search engine; it is a collaborator. To harness its full potential, you must engage with it as such. By giving it a voice, a character, and a purpose, it functions as intended—an expert without any clue of the outside world.
Go deeper: Want to teach your staff how to get consistent results with AI? Contact Todd Moses & Company.
