Your LLM needs Fine-tuning: What do I mean by this?
A field guide for non-builders who want smarter, specialized LLMs
You’re trying to use ChatGPT more seriously now.
You’ve built a prompt that works... kind of. But only when you copy-paste the full 8-line instruction set you wrote two months ago.
Maybe it’s something like this:
“You are a product marketing strategist. Your tone is concise, warm, and business-friendly. Your job is to rewrite these feature descriptions to make them clearer to non-technical customers. Focus on benefits, not features.”
You paste that into ChatGPT.
Then the feature descriptions, as detailed as possible (could be technical too).
Then add "Make it sound like Apple, not AWS." (because who wouldn’t want to sound as sophisticated as Apple does).
And still, half the time you get this from ChatGPT:
“Our platform enables seamless integration and provides users with robust capabilities across the board.”
Which is a fancy way of saying... nothing.
You know prompting helps. But it’s starting to feel like your AI needs to learn how you work, not just respond to what you type.
That’s when it’s time to consider fine-tuning.
🔧 What Fine-Tuning Really Means
Fine-tuning is a way of giving your AI a permanent memory of how you want it to behave… without having to prompt it every time.
It’s not about giving the model new knowledge. It’s about giving it your patterns.
How you answer support questions.
How you explain your product.
How your brand sounds when it writes anything.
It’s the difference between briefing a smart freelancer every time…
vs. training a teammate who just gets it.
🧠 When does Fine-Tuning Make Sense
You probably don’t need it for creative writing, exploration, or general use.
But you might want to fine-tune when:
You're doing the same task repeatedly with slightly different inputs
You need a reliable tone, structure, or formatting without re-explaining it
You’re building internal tools, workflows, or assistants that others will use
If you’ve ever thought, “Why do I need to keep reminding ChatGPT how to sound like us?” → This is for you.
🧰 How to Prep for Fine-Tuning (Even Without a Developer)
Before you fine-tune, you need examples. Not technical ones, just real ones.
Here’s how to start:
Gather your most common input–output pairs. Think: support questions and answers, sales objections and replies, marketing copy and rewrites.
Collect writing that reflects your tone. Emails, blog posts, call notes or anything with your team’s natural voice (every person and company has a unique way of communicating information or depicting something, consciously or subconsciously)
Identify formatting rules. Do you always use bullet points? A TL;DR? A problem–solution–benefit flow?
Flag common failure cases, things the model often gets wrong or misinterprets. (Like it can be too casual or too professional, it writes long paragraphs when you want short points, etc.)
Even 10–20 examples can go a long way.
You don’t need to format anything special. Just put each “input” and “ideal output” in a doc… like a dialogue. This becomes your training set.
🧠 What to Do With That Prep Material
Once you’ve collected the data, here are your next steps:
If you’re using ChatGPT Pro, you can fine-tune GPT-3.5 through OpenAI's interface (via the playground or API). You’ll need a .jsonl file, but you can use tools like OpenPipe or FineTune GPT to help format and upload your data. (However, using playground in OpenAI is not free)
If you're not ready to fine-tune yet on the OpenAI Playground, you can “simulate” it by:
Turning your examples into reusable prompt templates - you can literally just copy-paste all the examples in ChatGPT and ask it to extract the “tone guide” or a “writing style” based on how you want a typical output to look like
Using ChatGPT’s Custom Instructions feature to paste in your tone guide, writing style, and ideal responses.
Storing your training examples in a Notion or Google Doc, uploading the doc into the “Add Files” option in the chat, and prompting with:
“Based on the examples in this doc, answer the following question…”
It’s not true fine-tuning… but it’s often good enough to get 80% of the benefit with 0% of the complexity.
🔁 What Fine-Tuning Feels Like (Side-by-Side)
Let’s say you're answering product questions in a support flow.
Without fine-tuning:
“What does the analytics dashboard do?”
Response: “The analytics dashboard offers a variety of features that help you monitor performance.”
With fine-tuning:
“What does the analytics dashboard do?”
Response: “It shows you which pages users visit most, where they drop off, and how long they spend inside your product — all in one place.”
Same question. Different outcome.
You didn’t change the prompt… you changed the model.
⚖️ Fine-Tuning Isn’t for Everything
To be clear: fine-tuning has limits.
It doesn’t fetch fresh data. That’s something that RAG does. (We’ll discuss and explore RAG in my next post)
It doesn’t generate wildly creative ideas… prompting does that.
But if your model is giving half-right answers for things your team has already figured out, fine-tuning is the missing link.
🧪 Try This in ChatGPT
Here’s a quick test to see if your use case is ready for fine-tuning. If you’re looking for ChatGPT support as a sales rep, type the following:
“You’re a sales rep at our company. A prospect asks, ‘How is your pricing structured?’ Answer like we would on a live call.”
Then follow up with:
“What information are you missing to answer this properly?”
If the model says things like:
“Details on pricing tiers”
“How competitors compare”
“What pain points we typically address”
... and suppose you’ve already answered those in dozens of docs, call scripts, or FAQs, then you’re ready.
PS. Do comment if you want a guide like this built out for other specific business functions as well. Happy to create one :)
✍️ Final Thought: Fine-Tuning Isn’t About Smarter AI. It’s About Familiar AI.
It doesn’t just save time. It saves trust.
Instead of managing longer and longer prompts, you build a model that thinks like your team.
You don’t need to be a developer to benefit from it. You just need to start acting like a teacher.
Next up on LLMentary:
We’ve explored prompting and Fine-Tuning so far. Next up, we’ll dive into Retrieval-Augmented Generation (RAG). We’ll see what it is, why it’s so crucial to improving the quality of AI outputs, and how it helps your AI stop guessing… and start looking things up.
Subscribe to get the latest info chunks right to your inbox. I send 3 new weekly topics you need to understand to fully appreciate the power of LLMs!
Until then, stay curious.
Share this with anyone you think will benefit from what you’re reading here. The mission of LLMentary is to help individuals reach their full potential. So help us achieve that mission! :)