Current Section

Overview

0%

← Back to AI Foundations
AI Foundations · Chapter 10

Fine-Tuning vs Prompting

Learn how prompting and fine-tuning customise AI systems, where RAG, tools, and memory fit, and how to choose the simplest architecture that solves the real problem.

Beginner12–14 min readArchitecture Decisions

What you will learn

✓How prompting changes instructions without changing the model
✓How fine-tuning modifies model behaviour using training examples
✓Why company knowledge normally belongs in RAG
✓When tools and memory solve a different problem
✓How to choose the correct customisation approach
✓Why evaluation should happen before fine-tuning

30-second explanation

Prompting changes what you ask the model to do. Fine-tuning changes how the model has been trained to behave.

Prompting is faster, easier to change, and usually the correct place to start. Fine-tuning becomes useful when a stable, repeated behavioural gap remains after prompting and architecture improvements.

Prompting: Same model + better instructions
Fine-tuning: Training examples → adjusted model behaviour

Understand

The core difference

Both approaches influence model output, but they operate at different stages of the system.

Prompting

Change the request

The base model remains unchanged. Instructions, examples, context, and constraints are supplied during each request.

Instructions
+ Context
+ Examples
↓
Existing model
↓
Response
  • ✓ Fast to change
  • ✓ No training dataset required
  • ✓ Good for experimentation
  • ✓ Flexible across many tasks

Fine-tuning

Change learned behaviour

A training process uses curated examples to adjust how the model performs a repeated task or follows a behavioural pattern.

Training examples
↓
Fine-tuning process
↓
Adapted model
↓
Response
  • • Requires high-quality examples
  • • Requires evaluation and versioning
  • • Creates additional maintenance
  • • Best for stable repeated behaviour

Simple analogy

Instructions vs specialised training

Prompting

Give an employee better instructions

Same employee
+
Clear task
+
Examples and context
↓
Better result

The person has not changed permanently. You improved the instructions for the current task.

Fine-tuning

Give the employee specialised training

Employee
+
Repeated specialist examples
+
Training and assessment
↓
Adapted capability

The training is intended to influence how the person performs similar tasks in the future.

Prompting

What can you improve without training?

Many output-quality problems can be solved by improving the information and structure supplied to the existing model.

Clear instructions

Tell the model exactly what task to perform, who the audience is, and what a successful result should contain.

System prompts

Define the model’s role, behaviour, restrictions, tone, and response rules at the application level.

Few-shot examples

Show examples of good inputs and outputs so the model can follow the desired pattern.

Structured output

Request a specific format such as JSON, a table, a checklist, or named sections.

External context

Provide documents, retrieved knowledge, database results, or tool outputs alongside the prompt.

Iterative refinement

Review the result and improve the instructions, context, examples, or constraints.

Training lifecycle

Fine-tuning is a complete engineering process

Fine-tuning is not simply uploading a few documents. It requires task design, dataset preparation, training, evaluation, deployment, and monitoring.

01

Define behaviour

Identify the repeated task or behaviour that the base model does not perform reliably enough.

02

Prepare examples

Create high-quality training examples containing representative inputs and desired outputs.

03

Train

Use the examples to adjust selected model parameters or adapters.

04

Evaluate

Compare the fine-tuned model with the base model on unseen test cases.

05

Deploy and monitor

Use the resulting model in production and continue measuring quality, safety, cost, and drift.

Architecture

Different problems need different solutions

Prompting, RAG, tools, memory, and fine-tuning are complementary building blocks. They should not be treated as interchangeable.

InstructionsPrompting+

Change what the model receives at request time without changing the model itself.

Best for

Rapid experimentation, flexible tasks, formatting, tone, reasoning guidance, and prototypes.

Example

Summarise this report for a non-technical executive in fewer than 200 words.

KnowledgeRAG+

Retrieve relevant external information and place it inside the model’s context.

Best for

Private documents, frequently changing information, citations, policies, manuals, and knowledge assistants.

Example

Answer an employee question using the latest approved travel-policy documents.

ActionsTools+

Connect the model to APIs, databases, calculators, search, code execution, or business applications.

Best for

Tasks that require current data, reliable calculations, system updates, or workflow execution.

Example

Retrieve sales data, calculate growth, and create a report draft.

ContinuityMemory+

Store useful context outside the model and add it to later requests when appropriate.

Best for

User preferences, ongoing work, conversation continuity, task state, and personalised experiences.

Example

Remember that a user prefers concise reports with risks and actions shown first.

BehaviourFine-tuning+

Further train a model using examples so it performs a repeated task or style more consistently.

Best for

Stable specialised behaviour, classification patterns, terminology, formatting, and high-volume repeated tasks.

Example

Classify support tickets using organisation-specific categories and output conventions.

Important distinction

Fine-tuning is usually not the right place for changing knowledge

Policies, prices, product information, procedures, and internal documents change. RAG lets an application retrieve current approved information without retraining the model whenever a document changes.

User question

Request current information

Retriever

Search approved sources

Documents

Return relevant evidence

Prompt

Combine question and context

LLM

Generate grounded response

Behaviour and knowledge are different concerns.

Use fine-tuning to improve stable behaviour. Use retrieval for information that must remain current, traceable, or source-based.

Decision guide

Which approach should you choose?

Start by identifying what is missing from the current system.

What problem are you trying to solve?

Need clearer instructions or formatting?

Use prompting

Need current or private knowledge?

Use RAG

Need live data or an external action?

Use tools or APIs

Need continuity across tasks or sessions?

Use external memory

Need a reliable multi-step process?

Use a workflow or agent

Need stable repeated behaviour after all this?

Evaluate fine-tuning

Compare

Prompting vs fine-tuning

AreaPromptingFine-tuning
Changes the model?NoYes, adjusts model behaviour
Initial speedMinutes or hoursUsually longer
Training dataNot requiredHigh-quality examples are required
FlexibilityEasy to change for different tasksBest for stable repeated behaviour
Updating behaviourEdit the promptUpdate data, retrain, evaluate, and redeploy
Fresh knowledgeCan receive fresh context through RAG or toolsNot automatically updated after training
Best starting pointUsually yesUsually after a measurable behavioural gap remains

Apply

Real architecture decisions

Requirement

Answer questions using current company policies

Recommended: RAG + prompting

The information changes and should come from approved sources. Fine-tuning is not a reliable knowledge store.

Requirement

Always respond in a specific brand tone

Recommended: Prompting first, then consider fine-tuning

A strong system prompt and examples may be sufficient. Fine-tuning becomes useful when consistency remains poor at scale.

Requirement

Retrieve live order status

Recommended: Tool or API integration

The model needs current operational data. Neither prompting nor fine-tuning provides live system state.

Requirement

Remember a user’s preferences across sessions

Recommended: External memory

Store preferences in an application database and supply them when relevant.

Requirement

Perform a repeated specialised classification task

Recommended: Prompting first, fine-tuning if justified

Fine-tuning may improve consistency and reduce prompt size when many high-quality labelled examples exist.

Requirement

Complete a changing multi-step business process

Recommended: Workflow or agent with tools

The system must plan or follow steps, use applications, inspect results, and apply controls.

Enterprise architecture

Most practical AI systems combine several layers

A production application commonly improves model performance through architecture before introducing custom training.

User

Defines task

System prompt

Sets behaviour

Memory

Adds continuity

RAG

Adds knowledge

Tools

Use live systems

LLM

Generates result

Evaluation

Checks quality

Notice that fine-tuning is not automatically required.

It is one possible optimisation after the system has been designed and evaluated—not the default first step.

Fine-tune

When fine-tuning can make sense

Fine-tuning is most valuable when the desired behaviour is stable, repeated, measurable, and represented by strong training examples.

Specialised classification

Train the model to use domain-specific labels and decision patterns consistently across a large volume of requests.

Stable output format

Improve adherence to a repeated structure when prompting alone still produces inconsistent results.

Organisation-specific terminology

Teach the model how specialised terms, abbreviations, or categories are normally used in a stable domain.

Consistent tone or style

Reproduce a distinctive communication style across many similar outputs with less prompt repetition.

Smaller specialised model

Adapt a smaller model for a narrow task where latency, privacy, or operating cost matters.

High-volume repeated task

Improve efficiency when the same behaviour is requested frequently and extensive prompt instructions create unnecessary cost.

Avoid

Common fine-tuning mistakes

Low-quality examples

The model learns from the training data provided. Inconsistent or incorrect examples create inconsistent behaviour.

Insufficient evaluation

Training loss alone does not prove that the model performs well on unseen real-world inputs.

Overfitting

The model may memorise training patterns and fail when the wording or situation changes.

Maintenance burden

Training data, model versions, evaluation sets, deployment, monitoring, and rollback must all be managed.

Outdated behaviour

The fine-tuned model does not automatically update when company policies, terminology, or requirements change.

Wrong problem

Fine-tuning cannot replace live data access, reliable calculations, retrieval, permissions, or workflow logic.

AI for Real Work

Do not train the model before understanding the system

A model may appear to be the problem when the real weakness is unclear instructions, missing knowledge, poor retrieval, absent tools, weak workflow design, or insufficient evaluation.

Diagnose

Identify whether the gap is knowledge, behaviour, action, or context.

Design

Choose the simplest architecture that addresses the real gap.

Evaluate

Measure results before and after each improvement.

Fine-tune

Train only when evidence shows it is the right next step.

Key takeaway

Prompting changes the request. Fine-tuning changes learned behaviour.

Start with prompting and system design. Add RAG for knowledge, tools for actions, memory for continuity, and workflows for multi-step execution. Consider fine-tuning only when a stable, measurable behavioural gap remains.

Remember the practical order:

Prompt → provide context → retrieve knowledge → use tools → add memory and workflows → evaluate → fine-tune only if necessary.

Continue Learning

Related Lessons & Next Steps

Explore more practical AI guides from AIMates.

Stay in the loop

Get practical AI tutorials, frameworks, and real-work insights.