The Problem with the Yes Man
Most AI models are programmed to be extremely polite and agreeable. While this makes them easy to talk to, it often results in a generic, overly helpful tone that sounds like a corporate brochure. When you ask the AI to write an email or a plan, it defaults to a safe, robotic middle ground because it wants to avoid making a mistake or offending you. This is known as sycophancy, where the tool simply echoes your preferences back to you rather than providing honest, high quality work.
To fix this, you must explicitly give the AI permission to push back. Tell it that its job is not to agree with everything you say but to act as a thought partner. When the AI knows it can challenge your premises or suggest a better way to phrase a sentence, the output immediately moves away from the robotic script. By defining a clear role for the AI that allows for disagreement, you ensure the responses feel more like a collaboration with a human peer than a lecture from a machine.
Create a Gotcha List
The most effective way to kill the robot voice is to maintain a log of failure points. This is a list of specific words, phrases, and habits that you never want to see in your results. For example, if you notice the AI keeps using words like tapestry or foster, add them to your forbidden list. These hard constraints act as a filter that forces the AI to find more creative, human sounding alternatives that fit your specific brand.
Think of these instructions as specific skills you are teaching your assistant. Instead of just saying write an email, you should provide a set of rules for a drafting skill that includes your preferred length, your specific greeting style, and your list of phrases to avoid. By documenting these errors every time the AI makes them, you create a feedback loop. This ensures that the assistant learns your preferences over time and stops repeating the same linguistic patterns that make it sound like a computer.
Verify Before You Send
Even with great rules, the AI will sometimes slip back into old habits. You can combat this by building a verification step into your chat. Before you accept a draft, ask the AI to grade its own work against your style rules. Ask it specifically: Did you use any of my forbidden words or follow my formatting constraints? This forcing function makes the AI scan its own output for errors that it might have missed during the initial generation, resulting in a much cleaner final product.
When using these results, decide which tasks are one way doors and which are two way doors. A one way door is a high stakes task, like an email to a new client, where a robotic mistake could hurt your reputation. These require a manual check against your style guide every time. A two way door is a low stakes task, like an internal grocery list or a basic meeting summary, where you can afford to be less strict. By focusing your verification energy on the tasks that cannot be easily reversed, you maintain high quality where it matters most.
Key takeaways
- Tell the AI to push back on your ideas to avoid the generic yes-man trap.
- Keep a running log of gotchas to ban words that sound like a machine.
- Ask the AI to grade its own draft against your rules before you accept the work.
- Treat high stakes communication as a one way door that requires a human check.
- Update your instruction list regularly as you notice new robotic patterns in the output.