Two Ways to Hit the Target
the two fundamentally different ways to get a language model to do what you want: adjusting your instructions, or retraining the model itself.
Part 1
Part 2
What the Sights Actually Adjust
a closer look at what prompting really changes about a model's behavior, and why it works without any retraining at all.
Part 3
Before the Archer Could Adjust Anything
how systems were customized for new tasks before pretrained language models made either prompting or fine-tuning genuinely practical options.
Part 4
Describing the Shot in Words
how far a well-crafted prompt alone can actually go in steering a model's behavior, and why it's usually the right first lever to reach for.
Part 5
When the Sights Run Out of Adjustment
the genuine signs that a problem has outgrown prompting, and needs the deeper, more permanent change fine-tuning provides.
Part 6
Retraining the Muscle Memory
what fine-tuning actually does to a model's weights, and why this change is deeper and more permanent than anything a prompt can achieve.
Part 7
The Archer's Practice Set
why the quality and relevance of a fine-tuning dataset matters more than almost anything else in determining whether the result actually improves the model.
Part 8
How Many Arrows Does It Take
the tradeoffs between fine-tuning dataset size and quality, and why more examples eventually stop helping.
Part 9
A Full Rebuild vs. a Small Adjustment
the difference between fully retraining every part of a model and adjusting only a small, targeted set of parameters.
Part 10
The Cost of a New Stance
a direct comparison of what prompting and fine-tuning actually cost, in both upfront effort and ongoing operation.
Part 11
When the Target Moves
why a fine-tuned model's advantage can decay over time as the underlying task or data drifts, and what that means for maintaining it.
Part 12
A Stance That Only Works for One Target
the overfitting risk in fine-tuning, where a model learns the quirks of its training data instead of the underlying task it was meant to learn.
Part 13
Sights That Work for Any Target
why prompting's flexibility is a genuine structural advantage over fine-tuning, not just a matter of lower upfront cost.
Part 14
Training the Stance, Then Adjusting the Sights
why fine-tuning and prompting aren't actually mutually exclusive, and how combining them often outperforms relying on either alone.
Part 15
Testing the Archer Before the Competition
why evaluating a fine-tuned or prompted model requires the same rigor, on the same held-out data, before either goes into production.
Part 16
When the Archer Forgets the Old Targets
catastrophic forgetting: the risk that fine-tuning for a new capability quietly degrades capabilities the model already had.
Part 17
The Coach Reviewing Form
how instruction tuning and reinforcement learning from human feedback are themselves a form of fine-tuning, applied to make base models genuinely usable.
Part 18
Choosing the Right Approach for the Competition
a practical decision framework that pulls together every factor this series has covered, for choosing between prompting and fine-tuning on a real project.
Part 19
The Economics of Training a New Archer
the organizational-level costs of maintaining fine-tuned models over time — infrastructure, retraining cycles, and specialized skill — beyond any single project's calculation.
Part 20
The Archer's Full Toolkit
reassembling every lever this series has covered into one complete picture of how to steer, retrain, and maintain a language model deliberately.