AI Content Generation - Q&A

[aplume612]

Hey All,

I don't really post on BHW (as you can see), and I feel a bit guilty. Lately, I've found a few members who have really helped me with domains/backlinks, thus, I'd like to provide some help to the community in return.

I do quite a bit of AI content generation. Like a lot - in excess of 100 million words, if I had to guess, since OpenAI's initial release. My experience with content generation covers most commercially available AI writing tools, and specifically, OpenAI and Anthropic APIs. Occasionally, I'll use ChatGPT and Claude, but emphasis on "occasionally".

A day or so ago I came across this BHW thread which was the catalyst for this thread. Total transparency here, I have no clue what else is on BHW pertaining to this topic, so forgive if I am repeating known information. Regardless, I'm sure I can add some value here, and happy to do so.

I'll start with some super general info (may or may not be common knowledge), if anyone has specific questions, feel free to post them here, or shoot me a DM, I'll do my best to provide a timely response.

DISCLAIMER: I do not represent, work for, affiliate, and or promote any products I am mentioning on this thread. This information is solely based on my own personal use and experience with these products. If you don't like a product I mention, or you have a better way of using said product, feel free to say so/correct me, etc. I am not promoting any coding services, if anyone needs help, I'm happy to provide my own code as needed and or descriptions of what not/to do. When I refer to OpenAI/Anthropic, I am referencing my experience with their respective APIs. When I refer to ChatGPT/Claude, I am referencing my experience with OpenAI/Anthropic's commercially available paid products that most are familiar with using (ChatGPT and Claude). I have tested and used many other paid/premium AI services - I primarily use OpenAI/Anthropic.

Common AI Content Generation Knowledge/Practices/Pointers/Etc.

1. Commercially available AI content generation - in short, and IMO, they mostly suck. I'm not trying to make enemies here - just sharing on my own experience. I have thoroughly tested the top 4 with premium subscriptions (not free trials), and I've tested maybe another 7-10 that no one has ever heard of - OpenAI/Anthropic APIs produce significantly better results.
2. Generating content via API is night and day difference compared to using frontend ChatGPT/Claude - APIs produce far superior content.
3. I primarily use the APIs with Python - I use my preferred IDE about 60% of the time, and I use Anvil[.]Works the other 40% of the time. Anvil[.]Works is great if you want to make a program with a nice GUI. If you haven't done so already, it's a little tricky calling the OpenAI and Anthropic APIs within Anvil[.]Works, so if this is of interest to you, I'm happy to provide setup instructions.
4. If you are using ChatGPT or Claude, you should always begin with a "You are an expert..." prompt. For those of you who don't know what this is, you are basically telling the AI to act as an expert in a specific field. For example, you can start a prompt like, "You are an expert in SEO content writing. You are will write original, engaging content." > allow AI to respond > provide specific article instructions.
5. A huge issue everyone runs into is the verbose language that AI tends to use. Words and phrases such as bustling, aftermath, navigate, etc - the list goes on forever. The solution to this is a bit abstract and really depends on what you're trying to produce. Contrary to the BHW Thread I mentioned earlier, I find it's best to provide a vague instruction on word choice rather than listing out specific words to avoid - "Write at a nth-grade reading level. Use clear, authoritative tone." (change "nth" to whatever grade-level you are seeking)
6. You can in fact use templates to produce consistent outputs. Depending on which models you choose to use in your API calls, you can feed a "System Instruction" and a "Prompt" to the API. I've attached one of each as examples. I have NOT tried these in ChatGPT or Claude, so I have no clue what they will produce.
7. Per #6, I do prefer using the AI models that accept both a "System Instruction" and a "Prompt".
8. AI Generated Content With Human Revision - Depending on the content's purpose and your goals, you can use either API to produce 20k-30k words (or more) in a single output. I then use Clearscope to revise and optimize the content. I've slowly moved away from this method, but not completely.
9. 100% AI Generated Content - This is highly dependent upon content length. If you are shooting for maybe 2,000 words or less, I suggest using a single API call with solid System Instruction and Prompt. If you are seeking longer-form content, the best results I produce come from daisy chaining the API outputs. I'll provide a short example below.
10. I run all API calls on max tokens for the given models.
11. I request all outputs in markdown syntax to maximize quality of content rather than structure (HTML, CSS, etc.)
12. Lastly, simplicity is key. Treat the AI like a 5 year old child - provide simple, clear, and concise instructions. The more complicated the instructions, the more likely you will confuse the AI.

API Output Daisy Chain Example

• Function1 > OpenAI > create an article outline > Function1Output
• Function1Output > (optional function to summarize Function1Output) > Function2 > Anthropic > SEO the outline > Function2Output
• Function2Output > (optional function to summarize Function2Output) > Function3 > OpenAI > generate content for outline item 1 > Function3Output
• Etc...

Really unsure what the average person here knows on this topic, so feel free to impart some knowledge on me, or ask a question. Something I didn't really mention is my use of Google's NLP API for entities and salience - happy to answer questions on this as well (I also use TextRazor API in conjunction). Looking forward to hearing what everyone's thoughts.

IMPORTANT NOTE ON ATTACHED TEMPLATES: I use the attached "Prompt" template example for a very specific type of short-form content - I highly doubt anyone here would be generating such content. Therefore, use this as a starting point for your own template/needs. The variables that look like, "[TOPIC]" or "[INSERT]" are for you to change/provide to the AI.

 

[pauliakas]

Do you use Perplexity?

 

[aplume612]

Do you use Perplexity?

I have "tried" it, but not enough to come to a fair conclusion. If you have some solid experience with it, I'm all ears.

 

[pauliakas]

Perplexity API not working well for me, Frontend working good, sometimes has hallucinations but overall great for the money. Has some restrictions on content, claude and openai writes, perplexity no on some content.

 

[aplume612]

Perplexity API not working well for me, Frontend working good, sometimes has hallucinations but overall great for the money. Has some restrictions on content, claude and openai writes, perplexity no on some content.

I'll check it out more thoroughly and report back. Give me a week or so. Have you tried using Perplexity's API in conjunction with OpenAI/Anthropic API? Would be curious to see if either one can consistently correct/revise Perplexity's output when needed.

 

[natsuko]

Totally agree about the APIs being far superior to front-end tools like ChatGPT or Claude for large-scale content generation. Your point about daisy-chaining outputs is spot on, breaking tasks into smaller, logical steps really improves output quality. How do you balance token usage and cost efficiency when working with high-volume content?

 

[aplume612]

Totally agree about the APIs being far superior to front-end tools like ChatGPT or Claude for large-scale content generation. Your point about daisy-chaining outputs is spot on, breaking tasks into smaller, logical steps really improves output quality. How do you balance token usage and cost efficiency when working with high-volume content?

Balancing token usage/cost efficiency: I had a feeling this question would arise , and I hate admitting this but, I literally pay zero attention to the cost - my tokens are always set to each model's maximum tokens allowed.

The content I produce is solely for mine and my partner's business. We are arguably located in the #1 most competitive local market in the US for our business niche. For context, we are in the legal field and our top 8 local competitors spend ~$300k/mo on Google Ads, alone. Thus, my only priority is producing the highest quality content possible, and as much as possible.

That said, when I first began messing with the AI APIs, I made a decent attempt at balancing the token/cost usage. Full disclosure, this was almost two of years ago, so this may not be true today, but I found that simply limiting the tokens setting produced inferior content. I ended up running max tokens, then managing token usage by specifying content length - this became the tricky part. I tried two separate ways:

1. Instructions/Prompt: I specified word count and or character count limits, but the results were fairly inconsistent for me - these results may suffice for someone else's needs, so I wouldn't discount it.
2. Daisy Chaining: I wrote intermediary functions that checked content length (not token usage) in between API calls. I would then feed the info to the API calls with counters. Each subsequent API call would receive this info (amount of content produced, and amount of content to be produced). This was the only method I tried within my daisy chains.

Ultimately, I completely eliminated any form of cost/token monitoring. I have not performed a formal comparison, but based on my observations, it "felt" as though I spent more time, and more money on tokens, trying to revise content produced with token usage in mind. As quality gradually became my priority, I kept making small tweaks to my chain which made it more efficient, and I believe it ultimately reduced costs in the long run.

On a side note, I am a huge believer in finding the best temperature, top_p, frequency, and presence, settings for your use case. I think this is another way of managing costs, sort of a derivative to token usage.

Based on your question, I'm assuming you have your own way of balancing tokens/cost usage - would love to hear about it if you don't mind sharing. I'm certainly not opposed to implementing something for cost/token management, I just have to maintain the existing quality.

 

[aplume612]

Totally agree about the APIs being far superior to front-end tools like ChatGPT or Claude for large-scale content generation. Your point about daisy-chaining outputs is spot on, breaking tasks into smaller, logical steps really improves output quality. How do you balance token usage and cost efficiency when working with high-volume content?

My apologies, I forgot one other fairly important aspect pertaining to token/cost usage management - the input and output syntax. For a long time, I was using JSON formatted instructions and prompts, and requesting HTML outputs. I didn't realize how many more tokens JSON inputs/outputs consume until year or so ago. I ended up switching to plain text inputs and markdown outputs. This made a noticeable difference in cost (not significant, but noticeable).

 

[itz_styx]

a while ago i wrote a post about prompt generation, maybe it gives you some ideas: https://www.blackhatworld.com/seo/h...-generic-as-the-usual-chatgpt-output.1535048/

 

[aplume612]

a while ago i wrote a post about prompt generation, maybe it gives you some ideas: https://www.blackhatworld.com/seo/h...-generic-as-the-usual-chatgpt-output.1535048/

Absolutely love your post! You made some phenomenal points:

1. "They mostly focus on the structure of the article, rather than the content itself."
2. "Well thats just simply basic instructions and far from advanced. the tone of the output will be generic."
3. "...davinci-003 model from OpenAI was much better than the chatGPT models..."

With regard to point 1 and 2, I totally agree - these are not what I consider to be "advanced" prompts. However, for menial tasks, I believe the AI provides a slightly better output when it's an "expert" in something. I have found great success in providing the AI a template for the structure, not so much explaining the structure. I then breakdown the writing style as you propose in your thread. Have you tried providing some type of template for structured outputs rather than explanations? If so, would love to know what happened - I never had much luck with structured outputs that came from explanations.

Regarding point 3, my best piece of content was generated on the last day of access to davinci-003, in the final hours - no clue why they deprecated it, but certainly wish they didn't. Which model do you prefer using, now? And, do you switch between models at each stage of your process?

 

[itz_styx]

Absolutely love your post! You made some phenomenal points:

1. "They mostly focus on the structure of the article, rather than the content itself."
2. "Well thats just simply basic instructions and far from advanced. the tone of the output will be generic."
3. "...davinci-003 model from OpenAI was much better than the chatGPT models..."

With regard to point 1 and 2, I totally agree - these are not what I consider to be "advanced" prompts. However, for menial tasks, I believe the AI provides a slightly better output when it's an "expert" in something. I have found great success in providing the AI a template for the structure, not so much explaining the structure. I then breakdown the writing style as you propose in your thread. Have you tried providing some type of template for structured outputs rather than explanations? If so, would love to know what happened - I never had much luck with structured outputs that came from explanations.

Regarding point 3, my best piece of content was generated on the last day of access to davinci-003, in the final hours - no clue why they deprecated it, but certainly wish they didn't. Which model do you prefer using, now? And, do you switch between models at each stage of your process?

well yes it helps a bit if you tell it to be an expert in something, but then again without specific instructions it wont help that much. so you want to give it linguistic rules so it wont sound so generic.
in my prompt templates i give it various linguistic instructions and the tone it should use, that really helps to get some creative output and far from what you usually see.

as for point 3. yea sadly they removed davinci and it was the best model for text generation, even better than its current GPT4o, but well we have to work with the current ones so i do use GPT-4o now and its not as bad as GPT-3 was for example, but still miss davinchi to be honest

 

[aplume612]

well yes it helps a bit if you tell it to be an expert in something, but then again without specific instructions it wont help that much. so you want to give it linguistic rules so it wont sound so generic.
in my prompt templates i give it various linguistic instructions and the tone it should use, that really helps to get some creative output and far from what you usually see.

as for point 3. yea sadly they removed davinci and it was the best model for text generation, even better than its current GPT4o, but well we have to work with the current ones so i do use GPT-4o now and its not as bad as GPT-3 was for example, but still miss davinchi to be honest

Totally, GPT-3 was crap, and that's when I started using Anthropic a bit more. Now I'm probably 50/50 between the two. Do you have any experience with Anthropic? I feel like it has more of a "natural flow" to it, sort of similar to the davinci-003 outputs.

 

[itz_styx]

Totally, GPT-3 was crap, and that's when I started using Anthropic a bit more. Now I'm probably 50/50 between the two. Do you have any experience with Anthropic? I feel like it has more of a "natural flow" to it, sort of similar to the davinci-003 outputs.

nope havent tried anthropic .. don't find enough time these days to keep up with all the new things that pop up in the AI world, but i will try to check it out does it have an API to automate ?

 

[aplume612]

nope havent tried anthropic .. don't find enough time these days to keep up with all the new things that pop up in the AI world, but i will try to check it out does it have an API to automate ?

Yes, and honestly, it's the closest thing I've found to a davinci-003 output that I can find. Don't get me wrong, it's not the same, but it's the closest I can find. Mixing the two (OpenAI/Anthropic) produce the best results for me.

 

[itz_styx]

Yes, and honestly, it's the closest thing I've found to a davinci-003 output that I can find. Don't get me wrong, it's not the same, but it's the closest I can find. Mixing the two (OpenAI/Anthropic) produce the best results for me.

sounds good then i guess i'll have to try it.

 

[aplume612]

sounds good then i guess i'll have to try it.

check this out, thought you might like it...

https://twitter.com/x/status/1864535178133250318