This is Lesson #23 of the ANTIghostwriter course — a free, complete system for creating authentic content with AI assistance.
New here? Start from the full course overview.
Previous lesson: #22: Automate Cross-Platform Posting with Hypefury
What You’ll Learn
AI makes predictable mistakes — and they’re all fixable. In this lesson, you’ll learn to identify and correct: posts exceeding character limits, language hallucinations (foreign words slipping through), made-up facts and fictional examples, article length inconsistencies, and “chat fatigue” from long context windows. These issues happen to everyone; now you’ll know how to handle them.
Time to complete: ~8 minutes to read (reference as needed)
Artificial intelligence remains just that—artificial intelligence. It’s not human, won’t make conclusions for you, or make cognitive edits. I want to draw your attention to potential errors that may occur during content generation and explain what to do about them.
Posts May Exceed the Specified Length
We have a strict requirement in the prompt to Claude for post length—280 characters. However, she tends to exceed this limitation, especially for multi-line posts and lists.
One technique that helps improve this situation is including an instruction that each post should be accompanied by a character count. In most cases this helps, but Claude often produces an incorrect count—she simply cannot accurately count the characters in her own writing.
This might be an error in my prompting that I’ll eventually resolve, but it’s more likely just a feature of Claude’s current version.
Most posts fit within the limit, but for those that don’t, you should do the following: in the next iteration, tell Claude that some posts exceed the character limit and ask her to fix them. She’ll apologize and correct only the posts that are too long.
When you write:
Some posts exceed the character limit. Fix it.This usually works when used after the posts have been written. Alternatively, you can manually edit the posts yourself before publishing.
Ideally, you could enter this request immediately after the posts are written. However, you may want to rewrite some posts entirely, so I don’t recommend immediately spending precious tokens that we pay for with this model. You might prefer to rewrite these posts yourself.
Keep in mind that posts can be longer than 280 characters—it’s ultimately at your discretion.
Language Hallucinations
If you’re writing in a language other than English, sometimes phrases in your original language might slip through. I’ve caught such glitches several times, so be careful when copying posts. I emphasize again that blindly copying and pasting directly into threads is extremely shortsighted and can backfire.
Made-up Facts
These are hallucinations, meaning the AI can invent content that wasn’t in your source text or create ideas, especially if your source article is incomplete. Since the prompt strictly defines the article length, the original material may be insufficient, causing the AI to expand by making things up.
The model can fabricate content based on its training data or simply hallucinate independently. I recommend reading the output article carefully to catch these issues.
In the prompt, I’ve included requirements to help eliminate these glitches by clearly stating not to make things up or add thoughts not present in the original text.
You need to understand how this works. If the source text is complete and contains sufficient material for structuring the article in the specified format, the AI handles it well. But if there isn’t enough suitable material, watch carefully for hallucinations.
You can work as an editor and ask the AI to replace specific paragraphs or supplement the article with your own material.
I often did this before refining my prompts; I would tell the model to “remove this paragraph” or “take these paragraphs from my new note.” Now the prompt is designed to use my notes as a basis for writing and adhere to the specified train of thought, so it usually works as intended.
If you find material that doesn’t match what you envisioned, use the chat to correct the AI’s error.
Article Length
At one point, Claude started producing very condensed articles much shorter than specified in the prompt. These looked more like single posts rather than complete articles of several thousand characters.
I refined the prompt so the length requirement appears in multiple places, and I now write instructions before generating the article to observe all length restrictions. This isn’t always necessary, but if you encounter this issue, try explicitly instructing the AI to adhere to the specified article length.
The latest version of the prompt seems to have resolved this issue. It may have been a temporary malfunction or an older model being used under the 3.7 version label. If such glitches appear, try resetting the context by creating a new chat and going through the article generation steps again.
It’s quite possible that the second attempt will work much better.
“Chat Fatigue”
Due to the increasing context window, AI may start responding more slowly or hallucinate more, forgetting initial instructions. In this case, simply copy the original prompt into a new chat and start fresh. Chats with notes (where we format them) will fill up especially quickly since they have a long context due to text files and your growing library.
Feedback on Errors
These seem to be all the errors and issues I encounter, but artificial intelligence remains a black box—unpredictable in many ways. If you notice something I haven’t listed here, please provide feedback so I can refine the corresponding prompts.
We first need to understand the cause of any new issue, as it could be due to model changes or require adjustments to our approach. Most likely, these can be fixed by adjusting the prompt.