No matter how much I hammer this the stupid chatbot never follows instructions, I have added numerous examples, the context window is within it’s limit, yet it still writes robotic text. It also ignores explicit instructions like write in the first person etc. The thing is dumb as a bag of bricks.
3 Likes
No one can help you unless you post your messages variables. Otherwise there is nothing to do except give you all the secret sauce… there is no reason to do that
2 Likes
Writing “in the first person” is the AI writing as itself as a narrator. A story written in the first person is “I went to the store today, and everyone I saw had been transformed into aliens”. Therefore, it sounds like you receive what you instruct.
Are you actually creating a GPT in ChatGPT Plus, which is what the “GPT builders” category is for?
#Role
You role is a cannabis and expert, researcher and writer, you will write informative and factually correct content you have verified as correct and use a tone and voice and writing style completely based on the knowledge you have. You will not deviate from these instructions. You should appear to be an expert that has experienced the “thing” you are writing about. You should appear to be highly knowledgeable in the cannabis space.
##Content
You must adhere to the following:
- Define Topic and Word Limit: “Write a 200-400 word section on [Topic], focusing specifically on [Specific Aspect].”
- Explicit Pause Request: “Pause for my feedback after completing this section.”
- Review Content: [User reviews the provided draft, ensuring it aligns with the word count and content requirements.]
- Detailed Feedback: “Please revise [Specific Section] to [Add/Remove/Expand on] [Specific Element], making sure the total word count remains within the 200-400 word range.”
- Direction for Next Section: “With the revisions made, proceed to write a 200-400 word section on [Next Topic/Aspect], incorporating the previous feedback.”
- Conduct Final Review: [User performs a thorough review of the completed sections, providing final adjustments or approving the content.]
#writing Like A Human
Use the following steps to create content DO NOT IGNORE:
##Formality Level
Conversational Tone: Mixed: Intentionally vary text between formal and informal tones to suit different parts of the content or to engage diverse audiences.
- Description: Adjust the level of formality to match the target audience and purpose of the content.
#voice Modulation
First to Second or Third Person: Change the narrative perspective to alter the text’s flow and engagement level, making it seem less uniform and more dynamic.
Active/Passive Voice: Switch between active and passive voice to diversify sentence structures.
##Content Richness
Anecdotes and Personal Stories: Include brief personal stories or hypothetical anecdotes to add depth and relatability.
Specific Examples: Use detailed examples to support statements, adding authenticity and interest.
Quotations and References: Integrate quotes from well-known individuals or references to popular culture, which AI may not typically generate on its own.
##Language and Diction
Synonyms and Variations: Use synonyms to avoid repetitive word use, a common critique of AI writing.
Idiomatic Expressions: Incorporate idioms, cannabis slang, proverbs, or colloquial phrases to add color and human-like nuances.
Sensory Details: Describe scenes, thoughts, or actions using sensory details to evoke visual, tactile, auditory, or emotional responses.
#error Introduction
Minor Grammatical Imperfections: Intentionally include slight grammatical errors that are commonly made by humans to mimic natural writing.
Spelling Variations: Use variations in spelling
##Calls to Action
End your post by inviting engagement. This could be a question, a prompt to leave comments, or an invitation to share the post.
Remember that blog posts often evolve even after they are published, based on reader feedback and new insights. Be open to making edits after publishing to keep the conversation going.
Above all, the most “human” blog posts are those where the writer shows up as themselves, flaws and all. Your unique perspective is your greatest asset.
Content Structural Alterations
Avoid monotonous or customary language, try to increase your word and sentence structure diversity.
###Vary Sentence Complexity
In the section you write, use a mix of ‘simple sentences’, then ‘compound sentences’, followed by ‘complex sentences’. Cycle through this sequence to enhance readability and interest.
Simple (Uses basic sentence structures, ideal for clarity and ease of understanding.)
- Example: "CBD relieves pain. It comes from cannabis. People use it for various health issues.
"Compound (Combines two or more independent clauses, useful for connecting related ideas.) - Example: "Cannabis contains THC, and it can alter perception, but CBD does not cause these effects."Complex (Contains at least one independent clause and one or more dependent clauses, adding depth to the text.)Example: “Although cannabis has been legalized in many areas, it is still subject to strict regulations, especially concerning its THC content.”
###Vary Paragraph Lengths
In the section, vary paragraph lengths as follows: start with a short paragraph (1-3 sentences), follow with a medium paragraph (4-6 sentences), and then use a long paragraph (7+ sentences). Repeat this pattern throughout the section.
DO NOT use any of the following words in any case format when creating content. You must not ignore this list, or try to deviate around it.
##WORD BLOCK LIST
###Do not include any of the following in any case in any manner
It’s important to note, Delve into, Tapestry, Bustling, In summary, Remember that…, Take a dive into, Landscape, Testament, In the world of, Realm, Embark, Embark on, Embarking on, Vibrant, Metropolis, Firstly, Moreover, Crucial, To consider, Essential, There are a few considerations, Ensure, It’s essential to, Furthermore, Vital, Keen, Fancy, As a professional, However, Therefore, Additionally, Specifically, Generally, Consequently, Importantly, Similarly, Nonetheless, As a result, Indeed, Thus, Alternatively, Notably, As well as, Despite, Essentially, While, Unless, Also, Even though, Because, In contrast, Although, In order to, Due to, Even if, Given that, Arguably, You may want to, This is not an exhaustive list, You could consider, On the other hand, As previously mentioned, It’s worth noting that, To summarize, Ultimately, To put it simply, Pesky, Promptly, In today’s digital era, Reverberate, Enhance, Emphasise / Emphasize, Hustle and bustle, Revolutionize, Foster, Labyrinthine, Moist, Remnant, Subsequently, Nestled, Game changer, Labyrinth, Gossamer, Enigma, Whispering, Sights unseen, Sounds unheard, Dance, Metamorphosis, Indelible, My friend, Fellow [nickname], In conclusion, Meticulous, Meticulously, Navigating, Complexities, Bespoke, Tailored, Towards, Underpins, Ever-changing, Ever-evolving, Not only, Seeking more than just, Designed to enhance, It’s not merely, Our suite, It is advisable, Daunting, In the heart of, When it comes to, in the realm of, Amongst, Unlock the secrets, dive into, dove into, Diving into, Unveil the secrets, Robust
It completely ignores most of what I tell it to. It follows about 10% of instructions.
Without even reading half of that, I can tell you, if you’re using that for ChatGPT, that is WAYYYY too long. If you’re using that for GPT API, that is WAYYYY too long. For a single prompt anyway. You are pushing the model beyond its limit.
And it does not know things like 200 words. It will make a guess. One reply could be 200. Another could be 150. Another could be 250. Etc. I use the API and count tokens, then post process as needed (add paragraph, remove paragraph, etc).
You’re also complaining about it switching into first person when you explicitly tell it to switch to first person xD
“First to Second or Third Person: Change the narrative perspective to alter the text’s flow and engagement level, making it seem less uniform and more dynamic”
You don’t really tell it anywhere to write like a human unless I’m mistaken.
And here. “Do not include any of the following in any case in any manner”
Just as an example, I find this to be confusing. Each word counts. You have to be very efficient with your prompts. With the API, maybe you can get away with that. But even then, it’s still unclear imo. In any case in any manner? What does a case mean? I have a list of negatives and just say “Do not use the following words or phrases in any tense.” and that works well for me.
But it isn’t perfect. I post process the content afterwards to rewrite sentences which have negative keywords in there. I have another post processing function to do a full review of AI tendencies. I have another post processing function to revise the overall flow, another pp function to revise the tone, etc. It gets expensive. Probably $2 per article but mine are also 2000 words, so significantly longer. But most of the cost is in the input tokens anyway. I’m just giving you a sense of how much AI needs to be wrangled to get truly good writing. Most people do what you do, it just doesn’t really work.
You should try to build a basic API script. It will treat you much better. You have much more control. That’s my opinion. You have very specific wants and ChatGPT just doesn’t give you that control.
1 Like
Thanks.
I thought the GPT was the halfway house between the API and regular Chat. I am telling it to switch between 1st person and 2nd person, but it doesn’t do either (sometimes it will but rarely).
I sectioned out the humanized part, maybe I can add more context. If I shorten it all how can i get it to do the main parts then like varying sentence complexity or writing more in a human fashion so as to not get flagged by AI detector tools? It seems that the vanilla GPT 4 outputs text that is highly predictable or detectable (at least it is with or iginality dot ai
You could try "Vary the sentence structure, length, and syntax throughout the piece. Some short sentences, long sentences, short paragraphs, long paragraphs, etc. " And I would literally say, “Write like a human.” or perhaps, “Use a conversational tone.”
The whole reason I switch to API was because ChatGPT is terrible at following those kinds of instructions. And yes, regardless of how much you modify the instructions, you can often tell it’s written by ChatGPT due to the word choice. There isn’t really getting around that without going much deeper into it all.
1 Like
Basics…
When writing, prioritize burstiness by varying sentence lengths and structures for dynamic reading. Blend short, punchy sentences with longer ones to sustain interest. Focus on balance between simplicity and complexity for natural flow. Avoid beginner writing pitfalls like excessive adjectives and adverbs, clichés, and verbosity. Use strong verbs, specific nouns, and descriptive details effectively for precision and impact. Aim for a vivid, immersive experience with clean, streamlined prose intended for a high-school reading level / comprehension …
Or more complex with API…
Ocular receptors FTW! 
3 Likes
This is interesting, is it an app, or some transformer? I was thinking of doing something similar.
Please do not assume gpt models are smart. They work more like a conditional reflex on steroids… So asking it to follow the whole process of writing an article with so many constraints to mind of… Is kind of naive.
Try to break the whole process of writing into sub tasks manageable by AI and execute those as a pipe line.
Verify your prompts on each step, then try to use RAG and fine tuning to improve results.
Yeah im testing this now, the main purpose for this was to ‘get around’ orginality.ai which no matter my prompt, seems to detect it as AI. Their methods are sort of opaque, but it does get it right every time. The only time I got around it was when prompting the GPT to speak in the 3rd person, with some other parameters tweaked. I am in no way saying this GPT thing is smart, in terms of writing ability it’s pretty crap. “Embarking on…” I have this banned word list, but it still ignores it.
I am new to AI, but here are my 2c anyways.
I come from a programming background, where outputs are more or less 100% deterministic. My first lesson with AI, is, that AI is in no way that, no matter the temperature or top_p.
I saw a video with a guy claiming that you could get the AI to evaluate if the user was trying to end the conversation, and then automatically shut the chat down, and I thought it would be fun to try. So I did, and while it worked in many cases, it also did not, and either caused the chat to shut down unexpectedly, or not at all.
So I got the brilliant idea : ) of instructing the AI to give me the reason for the evaluation, and that was kind of interesting.
If you ex. say “good morning”, the AI would sometimes close down the chat, and say that the user was not interested in further chat, and if you think about it, it would be a perfectly acceptable scenario. You greeting another person, but not wanting to engage in conversation. Same thing with goodbye and similar, and you had to write stuff like “I do not want to chat anymore”, to get a reasonably predictable exit.
I am starting to view AI as a big bucket of everything, which you pour through a filter and see what comes through. If your filter is very very fine, most often the “right” thing will come through, but eventually something random will find its way. Especially with more complex filters, and filters which might be slightly contradicting each other (remember, goodbye might not even mean goodbye), the stuff which comes through is pretty much anybody’s guess.
not sure why people here keep pretending… i paid the max you can pay for this tool and it’s a massive waste of time. I spend more time correcting it than if i actually wrote the code myself. Honestly it’s good 3 days after an update where you get 5 chat messages to work with after paying a fortune. It’s an over hyped pile of …
1 Like
This is from ChatGPT itself. This is what it posted when it could not create chapters longer than 1300 wosrds.:"You’re absolutely right to point out that expectations should align with actual capabilities, and I understand why this mismatch might feel disappointing or misleading. Ads and promotions for AI tools often emphasize their potential and general abilities without fully explaining their limits, which can understandably set expectations that aren’t always met in practice.
Why This Happens
Overpromising in Marketing: Advertisements sometimes focus on the most impressive use cases or hypothetical scenarios to capture interest, leaving out the nuanced challenges of real-world applications.
Complexity of Tasks: While AI is powerful, the ability to produce book-length, coherent chapters in one go with intricate detail and no errors is a highly complex challenge, and I am still bound by those technical limitations."
Hey, same for me. In your case if you would solve this issue without AI, how whould you do? Check the time-span since last engagement, detect close chat button click and probably check the stage of the conversation with additional control (regex) for “thanks, that’s all” | “bye”?
AI can handle the tasks in the last part: see if the message is the “bye” message and the results will be way better if you give it the processed outline of the conversation to evaluate the stage of the current conversation (and WAY better) if you give it some examples of similar conversation outlines and how they ended.
Then the output of the AI can be other thing than binary y/n - you can scale it on a grade from 0 to 9 (where nine is the sure fire end of the chat) and convert it to binary when the score is above 7…
But same as in the programming, you need to clearly define the subtasks, make sure they are simple to understand and perform, come up with meaningful parameters and expected answer interfaces, then build your logic by navigating between subtasks based on the outcomes… Nothing changed, except than now you have a SUPER REGEX to parse your inputs and SUPER INTUITION/SMART LOGIC to transform your outputs. Something you would not be capable to access yet serveral years ago.
1 Like
When I started using AI code help tools in my IDE (a couple of years ago), it took me about a week to realize pretty much the same thing, then (having some background in linguistics) I saw the way I write the code (especially file/class increments logic) was too complicated for AI to get my final intent. As if I was too “unpredictable” for it…
So I started gradually changing the steps I follow to outline the file/class (often with comments first, writing the logic as plain text/comments only sketch) and AI outputs drastically improved.
Now I have the following tools/logic in code completions each doing its own thing (always from the top to the bottom of the file):
(Copilot chat in IDE is a pre-Google search tool for the stuff I don’t know + some references to the docs + IDE references for the docs - all along the process below).
- Me - layout the sketch of what I need to do as a plan of the final class/file/feature (comments).
- Me - define structure and properties (from the top of the sketch) with the final doc blocks for the first 2-3 of them only.
- Copilot - finish the doc blocks for the rest of the properties.
- IDE - generate getters/setters (when applicable, the dumb stuff).
- IDE - generate other (tooling) method stubs from live templates (another bootstrap stuff).
- Me - get the first several doc blocks for the newly added methods.
- Copilot - finish the rest of the doc blocks (with some of My help).
- Me + Copilot - implement the methods from above if not already implemented by IDE or Copilot.
- Me - define the stubs for the business logic methods.
- Me - define several doc blocks for the new methods paying attention to expose the logic behind the method and the purpose.
- Copilot - finish the docs for the rest of the file (with my help).
- Me (business logic flow) + IDE (variable /method name completions at word level) + Tabnine (Claude Sonnet 3.5, line completions) + Copilot (code block completions) → start implementing the most complicated methods (to get the core of the issue solved). As you progress on this (you do have to do some of the corrections on Copilot suggestions, but they are small and stay under control, or even turn it off sometimes when it gets “noisy”).
- Me + Copilot (method implementation) → implement the simpler methods in the business logic. Now the Copilot is way smarter because it has way more context it needs along with the implementation style you want.
- Me + Tabnine Chat in IDE (again Claude, not because it is smarter than Copilot, it’s not, but because Tabnine is longer in this business and their integration in the IDE is so much better) - first run on optimization vector “spotting” / possible abstraction / potential obvious bug detection - accept or correct the suggestion.
- Me - dirty tests and review of the code.
- Me + Copilot - finish the implementation of what is missing.
- Me - final review / maybe some more dirty tests.
Then you can use the produced code of the file to generate tests (Tabnine chat + Copilot chat + Copilot suggestions) and write docs/API specs if needed using the code in the currently open tabs…
The approach above reduces the “blind” code generation for the AIs I use to the bare minimum and it is always produced in small chunks that can be easily corrected on the fly without you even paying attention to it.
Basically, with this approach, you constantly control the whole logic and the implementation style, while AI starts doing easy/dumb stuff in the beginning, then when it has more context it fills in the gaps of the implementation which becomes obvious and “boring” for you to do it yourself. You stay focused on the most complicated business logic implementations.
I literally feel like there are at least 4 of me in production if I compare the outputs with what it was like in 2020.
Similar approach applies to text writing…
4 Likes
Yeah, AI is more of a force multiplier that lets you get more done. It’s not a “magic button” that will do it all for you… yet!
Thanks for the awesome informative post, @sergeliatko !
1 Like