Two Ways to Get AI to Do What You Want
Imagine you have hired a new employee. They are smart, well-educated, and capable, but they know nothing about your specific business. You have two basic strategies for getting them up to speed.
Strategy one: Give them clear, detailed instructions every time they do a task. "When a customer asks about returns, first check if the item was purchased within 30 days. Then verify the receipt. Then offer either a refund or exchange..." You spell everything out each time.
Strategy two: Send them through a training program. Over several weeks, they study your company's products, policies, and culture. After training, they know how to handle situations without needing step-by-step instructions every time.
These two strategies map directly to the two main approaches for customizing AI behavior: prompting (strategy one) and fine-tuning (strategy two). Understanding the difference between them, and knowing when to use each, is one of the most practical skills for anyone working with AI today.
What Is Prompting?
Prompting is the act of giving an AI model instructions in plain language at the time you use it. Every time you type something into ChatGPT or a similar tool, you are prompting. But prompting can range from a simple question to an elaborate, carefully crafted set of instructions.
Basic Prompting
At its simplest, a prompt is just a question or request:
"Write a summary of this article."
The model uses its general training to figure out what you mean and responds accordingly. For many tasks, this is perfectly adequate.
Prompt Engineering
Prompt engineering is the practice of carefully designing prompts to get better results. It turns out that how you phrase your request can dramatically affect the quality of the response. Here are some key techniques:
Be specific about what you want. Instead of "Write something about marketing," try "Write a 500-word blog post about three email marketing strategies for small businesses with limited budgets." The more specific your instructions, the closer the output will match your expectations.
Provide context. Tell the model who the audience is, what tone to use, and what constraints apply. "You are writing for a technical audience of software engineers. Use precise terminology. Keep explanations concise." This framing helps the model calibrate its response.
Give examples of what good output looks like. If you want the model to write product descriptions in a specific style, show it two or three examples of product descriptions you like. The model will pick up on the patterns and follow them. This technique is called few-shot prompting, and it is remarkably effective.
Assign a role. Starting your prompt with "You are an experienced financial advisor..." or "Act as a senior software engineer reviewing code..." primes the model to respond from that perspective, drawing on the relevant patterns in its training data.
Structure your prompt. For complex tasks, organize your instructions with clear headings, numbered steps, and explicit formatting requirements. Treat it like writing a detailed brief for a freelancer — the more organized your instructions, the better the result.
The System Prompt
Many AI applications use a "system prompt" — a set of instructions that is always included before the user's message. The system prompt defines the model's personality, capabilities, constraints, and behavior guidelines. When you interact with a customer service chatbot that always introduces itself as "Aria" and never discusses topics outside of the company's products, that behavior is defined in the system prompt.
System prompts can be quite elaborate — sometimes thousands of words long — and they are one of the most important tools for controlling AI behavior in production applications.
What Is Fine-Tuning?
Fine-tuning takes a fundamentally different approach. Instead of giving the model instructions at runtime, you modify the model itself by training it on additional data.
Here is the basic process:
-
Start with a pre-trained model. This is a general-purpose model that has already been trained on a broad dataset. It knows a lot about a lot of things.
-
Prepare training examples. You assemble a dataset of examples that demonstrate the behavior you want. These typically take the form of input-output pairs: "Given this input, produce this output." For a customer service application, you might have hundreds or thousands of examples of customer questions paired with ideal responses.
-
Train the model on your examples. The model's internal parameters — the millions or billions of numbers that define its behavior — are adjusted slightly so that its outputs more closely match your training examples.
-
The result is a customized model. After fine-tuning, the model has internalized the patterns from your training data. It produces outputs that reflect your specific style, terminology, and preferences without needing explicit instructions every time.
Think of it as the difference between reading a recipe each time you cook (prompting) and actually learning to cook through practice (fine-tuning). The recipe approach works fine, but the experienced cook is faster, more consistent, and better at improvising when things do not go exactly as planned.
What Fine-Tuning Changes
Fine-tuning adjusts the model's behavior in several ways:
Style and tone. A model fine-tuned on formal legal writing will naturally produce formal, precise language. One fine-tuned on casual customer chat will be conversational and friendly.
Domain knowledge. While the base model has broad knowledge, fine-tuning can deepen its expertise in a specific area. A model fine-tuned on medical literature will be more fluent in medical terminology and concepts.
Task performance. For specific, well-defined tasks — like classifying support tickets, extracting information from invoices, or generating SQL queries — fine-tuning can significantly improve accuracy compared to prompting alone.
Output format. If you always need the model to respond in a specific JSON structure, or to follow a particular template, fine-tuning teaches it to do this automatically rather than requiring format instructions in every prompt.
When to Use Each Approach
This is the practical question everyone wants answered: when should you invest in fine-tuning, and when is prompting sufficient?
Use Prompting When:
You are prototyping or experimenting. When you are still figuring out what you want the AI to do, prompting gives you the flexibility to iterate quickly. You can change instructions instantly and see how the output changes, without the time and cost of retraining.
The task is straightforward. For common tasks like summarization, translation, simple analysis, or general Q&A, good prompting with a capable base model will usually get you 90 percent or more of the way to ideal output.
Your requirements change frequently. If your instructions or guidelines change often — say, weekly promotions for a retail company, or evolving editorial guidelines for a publication — prompting is much easier to update than retraining a model.
You do not have training data. Fine-tuning requires a dataset of high-quality examples. If you do not have this data and would need to create it from scratch, prompting is the faster path to getting started.
You are using a very large, capable model. The most powerful models are often so good at following prompts that fine-tuning offers diminishing returns. When the base model can already do what you need with the right instructions, fine-tuning adds cost without proportional benefit.
Use Fine-Tuning When:
Consistency is critical. If you need the model to respond in a very specific way every single time — the same tone, the same format, the same level of detail — fine-tuning produces more reliable consistency than prompting, where the model might occasionally drift from instructions.
You need to reduce costs. This might seem counterintuitive since fine-tuning itself costs money. But fine-tuned models can often work with shorter prompts because they have already internalized your instructions. If you are making millions of API calls per month, the savings from shorter prompts can outweigh the one-time cost of fine-tuning.
You have a specialized domain. For fields with specialized terminology and conventions — legal, medical, financial, scientific — fine-tuning helps the model speak the domain's language fluently. A model fine-tuned on radiology reports will produce more accurate and natural-sounding output than a general model prompted to write like a radiologist.
Speed matters. Because fine-tuned models need shorter prompts, they can respond faster. In real-time applications where latency matters, this speed improvement can be significant.
You have high-quality training data. Fine-tuning works best when you have a substantial dataset of high-quality examples. If you have hundreds of expert-written examples of ideal outputs for your specific task, fine-tuning can leverage that data effectively.
Few-Shot Learning: The Middle Ground
Between zero instructions (hoping the model figures it out) and fine-tuning (retraining the model), there is a powerful middle ground: few-shot learning.
Few-shot learning means including a small number of examples directly in your prompt. Instead of just describing what you want, you show the model what you want with two, three, or five examples.
For instance, instead of saying "Classify these customer emails as urgent or non-urgent," you might say:
Classify customer emails as urgent or non-urgent. Here are some examples:
Email: "Our entire system is down and we cannot process any orders." Classification: Urgent
Email: "I'd like to update our billing address for next month's invoice." Classification: Non-urgent
Email: "We're seeing intermittent errors that are affecting about 20% of our users." Classification: Urgent
Now classify this email: "Can you send me a copy of our contract?"
Few-shot learning is remarkably effective because it gives the model concrete patterns to follow. It combines the flexibility of prompting (you can change examples instantly) with some of the precision of fine-tuning (the model has clear demonstrations of what you want).
The downside is that examples take up space in the prompt, which uses tokens and costs money. For high-volume applications, those extra tokens add up.
LoRA and Efficient Fine-Tuning
Traditional fine-tuning adjusts all of a model's parameters, which requires enormous computational resources. A model with 70 billion parameters needs substantial GPU power to fine-tune, making it expensive and slow.
LoRA (Low-Rank Adaptation) is a technique that makes fine-tuning much more efficient and accessible. Instead of adjusting all 70 billion parameters, LoRA freezes the original model and adds a small number of new, trainable parameters — often less than 1 percent of the original. These new parameters learn the specific adjustments needed for your task.
Think of it like this: instead of rebuilding an entire house to add a home office, you build an addition. The main house stays the same, and the addition provides exactly the new capability you need. LoRA is the room addition of the AI world.
The practical benefits are significant:
Much cheaper. LoRA fine-tuning might cost hundreds of dollars instead of tens of thousands, making it accessible to small and medium businesses.
Much faster. What used to take days can often be done in hours.
Easier to manage. Because the base model is unchanged, you can have multiple LoRA adaptations for different purposes — one for customer service, one for technical writing, one for legal analysis — and switch between them easily. It is like having interchangeable lenses for a camera rather than needing a separate camera for each type of photography.
Less data required. LoRA can produce meaningful improvements with fewer training examples than full fine-tuning, sometimes as few as a hundred high-quality examples.
Other efficient fine-tuning techniques include QLoRA (which combines LoRA with quantization to reduce memory requirements further) and adapter methods that insert small trainable modules into specific layers of the model.
The Cost/Benefit Framework
Making the right choice between prompting and fine-tuning involves weighing several factors. Here is a practical decision framework:
Consider the Volume
How many requests will this AI handle? For low-volume tasks (a few hundred requests per month), prompting is almost always sufficient. The overhead of fine-tuning is not justified. For high-volume tasks (millions of requests per month), fine-tuning can reduce per-request costs significantly.
Consider the Stakes
What happens when the AI gets it wrong? For low-stakes tasks like drafting social media posts or summarizing articles, prompting provides enough control. For high-stakes tasks like medical triage, legal analysis, or financial recommendations, the greater consistency of a fine-tuned model may be worth the investment.
Consider Your Data
Do you have high-quality examples of ideal outputs? Fine-tuning is only as good as the data you train on. If your examples are inconsistent, poorly written, or unrepresentative, fine-tuning will bake those problems into the model. It is better to invest in prompting than to fine-tune on bad data.
Consider Your Timeline
Prompting can be set up in hours. Fine-tuning takes days to weeks when you factor in data preparation, training, evaluation, and iteration. If you need a solution quickly, start with prompting and consider fine-tuning later as an optimization.
Consider Maintenance
Prompts are easy to update — change the text and you are done. Fine-tuned models need to be retrained when requirements change, which takes time and money. If your requirements are evolving rapidly, the maintenance burden of fine-tuning may outweigh its benefits.
A Real-World Decision
To make this concrete, imagine you run a mid-size e-commerce company and want to use AI to respond to customer emails. Here is how you might think through the decision:
Start with prompting. Write a detailed system prompt that explains your company's products, policies, tone of voice, and response guidelines. Include a few examples of ideal responses. Deploy this and monitor quality.
Evaluate the results. After a few weeks, you notice that the AI handles 80 percent of emails well, but struggles with complex return situations and sometimes uses language that does not match your brand voice.
Try better prompting first. Add more specific instructions for return situations. Include more examples. See if the results improve.
Consider fine-tuning if prompting plateaus. If you have gathered hundreds of ideal response examples from your customer service team, and prompting improvements have plateaued, fine-tuning using LoRA could help the model internalize your brand voice and handle complex scenarios more reliably.
Keep monitoring. Whether you prompt or fine-tune, continuous monitoring is essential. Customer needs change, products change, policies change. Your AI needs to evolve alongside them.
This iterative approach — starting simple and adding complexity only when needed — is how most successful AI deployments work in practice. The organizations that jump straight to expensive fine-tuning without first exploring what good prompting can achieve often waste time and money.
The Blurring Line
It is worth noting that the boundary between prompting and fine-tuning is becoming less distinct. Several trends are contributing to this:
Longer context windows mean prompts can include more examples, bringing few-shot prompting closer to the effectiveness of fine-tuning for some tasks.
Better base models are increasingly capable of following complex instructions out of the box, reducing the need for fine-tuning.
New techniques like RLHF (Reinforcement Learning from Human Feedback) and constitutional AI represent approaches that sit somewhere between prompting and fine-tuning, shaping model behavior through feedback rather than explicit examples.
The practical takeaway is that prompting should be your default starting point. It is faster, cheaper, more flexible, and increasingly effective. Fine-tuning remains a valuable tool for specific situations, but it is an optimization, not a prerequisite. Master prompting first, and you will know when and why fine-tuning is worth the investment.
See This in the News
Prompting has become a skill in its own right, with practical techniques that anyone can use to get better results from AI. For a hands-on guide, read 50 ChatGPT Prompts to Save Time at Work on AIWire.