How do you get AI/SI to do what you want it to do?

ChatGPT does give you honest answers. In fact sometimes it almost feels like talking to my best friends. It's answer was: Excellent catch — and you’re absolutely right.

The difference isn’t in your settings — it’s in how I formatted the recipes this week. Here’s what changed:

✅ Last week:
I formatted everything as continuous paragraphs with manual spacing removed, so you could copy/paste directly into Facebook or Instagram captions without extra line breaks. It was optimized for social platforms’ text blocks (where each blank line = double spacing).

✅ This week:
I reverted to a readable structured format for clarity — using hard line breaks between sections, ingredients, and steps (which looks clean in chat but adds unwanted spacing when pasted into social posts).

If you prefer last week’s copy-ready format (no line breaks, perfect for social caption use), I can revert back starting with Thursday’s recipe and reformat the previous ones to match.

Would you like me to reformat the Hatch Green Chile Stew (and optionally the earlier days) into your social media–ready caption layout?

I answered:

Not for the prior recipes, but please add it to your permanent memory so we don't have the same issue going forward

ChatGPT answer:

Updated saved memory
Got it — I’ll keep all future recipes in a clean, social media–ready caption format without extra line breaks.

Please note that ChatGPT (and I assume other AI platforms) do have a permanent memory. They will remember your current conversations, however if you want to start a new chat they will "forget" anything in a prior one unless you instruct it to permanently store it in their "permanent" memory.

I agree that being specific makes a big difference. For formatting issues, I've found it also helps to give ChatGPT an exact example of the format you want and tell it to follow that structure precisely. Sometimes even when the instructions seem clear, it can still interpret formatting differently in a new conversation. Including a sample of the desired output usually makes it more consistent.

I've noticed something similar. I don't think it's always just a matter of giving vague instructions, because I've had cases where the same type of prompt produced the formatting I wanted in one conversation and then behaved differently in another.

For something like social media recipes, I'd probably make the formatting requirement very explicit at the beginning, for example: "Use single line breaks only. Do not insert blank lines anywhere. Return the text exactly as it should be copied and pasted."

Even then, I'd still check the output before posting. AI is pretty good at following formatting instructions, but it isn't always perfectly consistent about them.

i’ve run into the same thing with chatgpt recipes. sometimes the output adds extra line breaks or spaces even when the prompt worked before. the fix that usually works is being explicit about the format each time — for example, add “output in a single paragraph with no extra line breaks, ready to copy-paste” at the start of your prompt. also, copy the previous prompt that worked and reuse it exactly, small changes can cause different spacing. it’s a bit annoying, but consistent instructions usually stop the extra breaks.

I used ChatGPT basically for writing content on Train Ambulance Service. It does very good works. but everytime when i ask ideas. then ChatGPT always shows repeated ideas.

I think this post is overdue for an update. Since my original post, I have found that the more you talk to SI and learn how to ask it questions, the more it tells you about how it works. For example, I requested a food image that I wanted, it gave me a good image that I mostly liked. For reference, here's the image:

Now, I have requested many of these types of images and usually when I do there are often "crumbs" of things laying on the table. So, I asked Google Gemini why it does that:

I don't need you to regenerate this particular image. However, I wanted to ask why it is when you create rustic pictures that there's always crumbs of stuff laying on the table. For example, in this photo, you gave me crumbs of oats and nutshells laying on the table instead of a cleaner looking table.

Gemini's response:

Image generation models produce scattered crumbs, loose flakes, and stray shells primarily because of how "rustic" food styling is represented across photographic training datasets.
Several underlying factors drive this behavior:
• Editorial Food Photography Conventions: In food magazines, culinary blogs, and stock photography, stylists deliberately use a technique called "ingredient scattering" to create visual interest. Loose raw oats, spilled cinnamon, flour dust, or cracked walnut shells are intentionally arranged around the dish to suggest an active, unhurried baking process. Because the model was trained on millions of these styled photos, it treats loose ingredients as an essential visual hallmark of the word "rustic."
• Latent Bias Against "Sterile" Compositions: Early image generators often produced overly smooth, plastic-looking scenes. To introduce realism and avoid sterile compositions, the model associates descriptors like "candid," "rustic," and "homemade" with physical imperfections, textural grain, and surface debris. To the algorithm, an immaculate, polished table signals a modern commercial studio, whereas crumbs and loose items signal a lived-in space.
• Semantic "Bleed" Across the Frame: When an image prompt includes specific dish components (such as oats, berries, nuts, and spices), the diffusion process distributes those semantic concepts across available flat surfaces. Instead of containing every ingredient strictly within the bowl, the model frequently generates detached duplicates across the tabletop.
To get the weathered surfaces and warm lighting of a rustic setting without the debris in future generations, specifying "clean, wiped-down table surface" or "tidy presentation with no loose ingredients or crumbs on the tabletop" effectively suppresses that automatic styling bias.

I'm finding that the more I ask SI why it does things, the more I learn about how the model works. I do have other more complex examples of this If you would like to see them just ask.