Pre-training and Post-training
Prompt engineering means teaching a model what to do via the context input into the model, and it is not the same as training or finetuning.
A Common Claim About Teaching a Model
People who work with language models often use the words training, finetuning, and prompt engineering as if they were interchangeable. Sometimes that looks harmless, because all three involve teaching a model to do something useful. But the words point to different activities, and mixing them up leads to confused plans, confused conversations, and confused expectations about what a model has actually learned. This article starts with the most common version of the confusion and then builds a clear way to tell the activities apart.
What do you think happens?
You paste ten sample emails into a chat with a language model, then ask it to write a new email. It writes in the same style as your samples. Has the model been trained on your emails?
Reveal answer
No, the model was not trained. This is prompt engineering.
The source material says that feeding examples into a model so it imitates them is describing prompt engineering, not training. The examples shaped one response through the context input, and the model itself was not reshaped.
Why the Word Training Gets Misused
The most common misconception runs like this: if you show a model examples of what you want and it starts producing similar output, then you must have trained it. The reasoning feels natural. The model learned something from your material, so the word training seems to fit. The source material points out that this use of the word is not technically correct. Calling prompt engineering training, or calling it finetuning, is a common misuse of terms that can confuse plans and conversations.
Prompt Engineering as Context Input
Prompt engineering is the process of teaching a model what to do via the context input into the model. The key phrase is context input. Everything the model sees in the current request, such as instructions, examples, background text, or a story about a desired style, is part of that context. You shape the model's behavior by deciding what goes into that input.
A helpful way to picture this is a simple flow. Your examples enter the context, the model reads them alongside your request, and the output reflects what it read. The adjustment lasts only as long as that input is present. Once the request is over and the context is gone, nothing about the model itself has been reshaped by the process. The next request starts from the same model, with only the new context you provide.
The source describes a real version of this pattern. A writer fed twenty-five childhood journal entries into ChatGPT so that it could mimic her younger self. Feeding those entries in is prompt engineering, because the teaching happens through the context input. The material is assembled into the prompt, and that prompt is all the model receives for this request.
Finetuning and What Changes in the Model
Finetuning is the process of adapting a pre-trained model to a specific task or need, and is often performed by application developers.
The source gives an example of this use: an application developer might finetune an OpenAI model to adapt it to their needs. Here the developer is not just pasting a few examples into a request. They are carrying out a training process that changes the model for their purpose. The difference is not that one activity uses examples and the other does not. The difference is where the teaching lands. Prompt engineering teaches through the input for a single request, while finetuning teaches the model through training.
Post-training as the Stage After Pre-training
Post-training refers to the process of training a model after the pre-training phase. Conceptually, post-training and finetuning are the same and can be used interchangeably. The terms still carry different signals about who is doing the work and what goal they have in mind.
Post-training is usually the term when model developers do it. The source gives an example in which OpenAI might post-train a model to make it better at following instructions before releasing it. Finetuning is the term when application developers do it, as in the example of adapting a model to their own needs. A single model can pass through both stages: a model developer may post-train it, and an application developer may then finetune that already post-trained model.
The source notes that pre-training and post-training make up a spectrum, and that their processes and toolings are very similar. Because of this, the line between the stages is more about purpose and actor than about a sharp technical boundary. The diagram shows the usual order, but the words you choose should follow who is doing the work and why.
Who Usually Does Each Activity
| Activity | How the model is taught | Who typically performs it | Changes the model itself? |
|---|---|---|---|
| Post-training | A training process after pre-training | Model developers | Yes, through training |
| Finetuning | A training process adapting a pre-trained model to a specific task or need | Application developers | Yes, through training |
| Prompt engineering | Context input placed into a request | Whoever writes the input; the source does not assign a specific role | No, the model is unchanged |
Post-training and finetuning are conceptually the same process, differing mainly in actor and goal.
Classifying a Described Activity
When someone describes an activity, you can sort it with one main question: is the teaching carried by the context input of a request, or by a training process applied to the model? If the teaching happens only through the input, it is prompt engineering. If a training process changes the model after pre-training, the next question is who is doing it and for what purpose. A model developer's post-training and an application developer's finetuning are conceptually the same, so the deciding factor is the actor and the goal described.
Mistakes Beginners Make With These Terms
Common mistake Calling prompt engineering training
What it looks likeSaying the model was trained when you pasted examples into a chat request.
Why it failsThe teaching happened through the context input, and the model itself was not changed.
Do this insteadDescribe it as prompt engineering, and say the adjustment lasts only while the input is present.
Common mistake Judging the activity by the output alone
What it looks likeDeciding that a model was finetuned because its answers now match a style.
Why it failsSimilar output can come from either activity, so the output does not reveal how the model was taught.
Do this insteadAsk how the model was taught: through request input, or through a training process.
Common mistake Assuming post-training and finetuning are different processes
What it looks likeTreating post-training as a more advanced kind of training than finetuning.
Why it failsThe source states that the two are conceptually the same and differ mainly in who does the work and why.
Do this insteadUse post-training for model developers and finetuning for application developers, and recognize they are the same kind of process.
Common mistake Using the word training as a general synonym for any way of getting better output
Trying the Classification Yourself
A team pastes twenty examples of polite refund replies into a chat window and asks the model to answer new customer messages in the same tone. Another team runs a training process on a pre-trained model using their own company's support conversations so the model adapts to their product. Classify each activity as prompt engineering or finetuning, and name the question that decided your answer.
Hints
- Start with the flowchart question: is the teaching carried by the context input of a request?
- If the teaching happens through a training process applied to the model, ask who performs it and for what purpose.
- Remember that post-training and finetuning are conceptually the same; the actor and goal are what distinguish the terms.
Key Takeaways
- Prompt engineering teaches a model what to do through the context input of a request, and the model itself is not changed.
- Training is a process that changes the model; post-training happens after pre-training, and finetuning adapts a pre-trained model to a specific task or need.
- Post-training and finetuning are conceptually the same; the terms differ mainly in who performs the work and why, with model developers usually post-training and application developers usually finetuning.
- To classify an activity, ask first whether the teaching happens in the context input; if it does not, ask whether a training process was applied and who is doing it.