Training the Stance, Then Adjusting the Sights

Opening Scene

The most effective archers don’t choose between a well-trained stance and adjustable sights — they use both together. A solid trained foundation handles the bulk of the work reliably, while the sights make the final, situational correction for each specific shot. Fine-tuning and prompting work exactly this way in practice: not a strict either-or choice, but two levers that often combine into something better than either alone.

In Plain English

A common, effective pattern fine-tunes a model on a task’s core, stable behavior, then still uses prompting on top of that fine-tuned model to handle situational context, formatting instructions, or details that vary call to call. This combines fine-tuning’s persistent behavioral shift with prompting’s flexible, per-use steering, and connects directly to combining fine-tuning with the retrieval techniques covered in this content library’s dedicated RAG series, where fine-tuning shapes how a model uses retrieved context and prompting supplies that context itself.

The Old Way

Before this combined approach was widely recognized as a legitimate, effective pattern, prompting and fine-tuning were often framed as a strict either-or choice:

• Projects sometimes framed the decision as fine-tuning versus prompting, as if choosing one meant abandoning the other entirely.
• There wasn’t yet a well-established practice of using prompting deliberately on top of a fine-tuned model, rather than treating fine-tuning as a complete, standalone replacement for prompting.
• Combining fine-tuning with retrieval-augmented generation specifically wasn’t yet a widely recognized, deliberate architectural pattern.

Recognizing this combined approach as often the most effective option, rather than treating the two levers as mutually exclusive, reflects real, accumulated practical experience.

What’s Changing (and Why AI Is the Reason)

1. Practitioners increasingly combine fine-tuning for a task’s stable core behavior with prompting for situational, per-call flexibility, rather than treating the choice as strictly either-or.
2. This connects directly to combining fine-tuning with retrieval-augmented generation, covered in this content library’s dedicated RAG series, where fine-tuning and prompting handle genuinely complementary roles.
3. This combined pattern feeds directly into the decision framework in Article 18, where “both together” is recognized as a legitimate, often optimal, option alongside choosing either alone.

The Metaphor, Fully Extended

The Archer	Combining Fine-Tuning and Prompting
A trained stance handling the bulk of the work reliably	Fine-tuning shaping a model’s stable, core behavior
Sights making the final, situational correction for each shot	Prompting supplying situational context and instructions per call
Both together outperforming either alone	Fine-tuning and prompting combined outperforming either used alone
A skilled archer using every available tool deliberately	A well-designed system using both levers deliberately and together

Quick Recap

• Fine-tuning and prompting aren’t strictly mutually exclusive — combining them often outperforms relying on either alone.
• A common pattern fine-tunes for a task’s stable core behavior while using prompting for situational, per-call flexibility.
• This connects directly to combining fine-tuning with retrieval-augmented generation, covered in this content library’s dedicated RAG series.
• The decision framework in Article 18 treats “both together” as a legitimate, often optimal, option.

Where This Fits in the Series

Article 14 covered combining both levers deliberately. Article 15 turns to how you actually test whether either approach, or their combination, is genuinely working.