Most “humanizing” prompts are actually making your AI dumber and clogging up its memory with useless rules.
We have all seen it: the double em-dashes (–), the overuse of words like “delve” or “tapestry,” and that distinctively flat, robotic cadence that screams “I was written by a machine.” I recently found a discussion started by a thoughtful contributor on the ChatGPT Prompt Genius forum that addresses this exact headache. Instead of fighting the AI with long lists of banned words, this expert proposes a much simpler, cleaner solution that leverages the model’s own training against itself.
The Core Strategy: Semantic Avoidance
The original poster argues that trying to micro-manage the AI’s output with specific formatting rules is the wrong approach. They found that standard fixes, such as telling the bot “do not use em-dashes” or “replace dashes with commas,” are inefficient and fragile. These instructions often fail as the conversation gets longer, or they force the model to focus so hard on what not to do that the quality of the content suffers.
The solution presented by the author is surprisingly elegant. They suggest adding a single line to your ChatGPT Custom Instructions (Personalization):
“Avoid common LLM patterns and phrases.”
According to the creator, this broad instruction works better than specific rules because it covers the punctuation issues (like the double dash) and the vocabulary issues (like the robotic buzzwords) simultaneously.
💡 Why Negative Prompting Clutters the Context Window
The first major insight from this professional is that negative prompting is a hidden drain on resources. When you tell an AI “Do not include –,” you are adding a constraint that the model must constantly check against. The author notes that this “clutters the context window for longer sessions.”
Here is why that matters: Every time the AI generates a token, it looks back at its instructions. If your instructions are a laundry list of “Don’ts,” the model has to process those negatives at every step. This doesn’t just waste tokens; it can make the model more rigid. By removing the specific negative constraints and replacing them with a general directive, the expert suggests you free up the model’s “cognitive load” to focus on the actual content rather than syntax policing.
✅ The Problem with Explicit Replacement Rules
The second insight deals with the fragility of basic replacement commands. The original poster points out that instructions like “replace em dashes with ..” require “constant guidance and is generally forgotten within a few messages.”
This is a common frustration known as “instruction drift.” As a conversation progresses, the AI tends to revert to its baseline training, which is heavily biased toward those robotic patterns. Specific syntax rules are often the first things to be forgotten as the context window fills up. However, the author implies that a high-level stylistic instruction like “Avoid common LLM patterns” sticks better because it defines the persona rather than a specific rule. It shifts the entire vibe of the output, making it more robust against drift than a fragile formatting rule.
📌 Leaning Into Predictive Capabilities
The most fascinating part of this discovery is the theory behind why it works. The creator suggests that this method succeeds because it “leans into the predictive capabilities LLM are trained on, versus attempting to exclude or avoid specific things.”
Think about how these models are trained. They have read billions of words. They know what “LLM patterns” look like because they generate them statistically. They also know what “human writing” looks like. By explicitly telling the model to avoid the concept of “LLM patterns,” you are effectively asking it to switch tracks. You are telling it, “You know that statistical cluster of words you usually use? Don’t go there.” This prompts the model to access a different subset of its training data: the subset that sounds more natural and less predictive. It is a psychological trick applied to code!
How to Apply This Fix
If you want to test this innovator’s method, here is the simple workflow to add it to your setup:
1. Open Settings: Go to your ChatGPT settings menu.
2. Find Personalization: Look for the section labeled “Custom Instructions” or “Personalization.”
3. Input the Command: In the box that asks how you would like ChatGPT to respond, paste the magic phrase: “Avoid common LLM patterns and phrases.”
4. Save and Test: Start a new chat and ask for a creative writing piece. You should notice fewer double dashes and less robotic fluff immediately.
It is refreshing to see a solution that subtracts complexity rather than adding to it. If you want to see the original debate or share your own results, check the link below.
💡 FAQ & Troubleshooting
What is the best way to stop the AI from using em-dashes (–) and repetitive patterns?
Instead of prompting the model manually in every chat, add the specific line “Avoid common LLM patterns and phrases” to your ChatGPT Personalization settings (Custom Instructions). This efficiently removes double dashes and reduces the usage of commonly repeated words by leveraging the model’s predictive capabilities rather than relying on strict exclusion rules.
Why do negative prompts like “do not include dashes” often fail?
Negative prompts (e.g., “do not use –“) clutter the context window, which reduces performance during longer chat sessions. Furthermore, simple exclusion rules are less effective than positive instructions because they require the model to constantly check against a “ban list” rather than flowing naturally.
Why does the model forget instructions like “replace dashes with periods”?
Basic replacement commands generally require constant guidance. In standard chat sessions, the model tends to forget these specific formatting constraints after a few message exchanges. Using the personalization setting creates a persistent instruction that applies to every response automatically.
Best way to replace em-dashes (–) and other common LLM patterns.
byu/nickakio in