What the Sights Actually Adjust

Opening Scene

Adjusting a bow’s sights doesn’t change the archer’s muscles, training, or draw strength at all. It changes where the arrow is aimed, given exactly the same underlying archer. Prompting works on this same principle: it doesn’t change anything about the model’s underlying weights or trained capabilities. It changes what the model is being aimed at, using exactly the same underlying capabilities the model already has.

In Plain English

Prompting works because a large pretrained language model already has broad, latent capabilities baked into its weights from training. A well-crafted prompt doesn’t add new capability — it directs the model toward using the right existing capability for the task at hand, through instructions, examples, and context provided at the moment of use. This is why prompting can be so effective without any retraining: it’s steering, not teaching.

The Old Way

Before large pretrained models carried this much latent, general capability, steering a system’s behavior without retraining it wasn’t really an option:

Earlier, narrower machine learning models often had no broad latent capability to steer toward — they were trained for one specific task and couldn’t be redirected toward a different one through instructions alone.
Getting a system to do something new usually meant training it specifically for that new thing, since there was no large reservoir of general capability already present to draw on.
The idea of “steering” a fixed model toward different behaviors purely through input phrasing wasn’t yet a practical concept for most systems.

Prompting as a genuinely powerful lever emerged specifically once pretrained models grew large and broadly capable enough to have real latent capability worth steering.

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

Large language models increasingly carry enough broad, latent capability that a well-crafted prompt alone can unlock significantly different behaviors, without ever touching the model’s weights.
This connects directly to the prompt structure techniques covered in this content library’s dedicated prompt engineering series — instructions, examples, and formatting are the concrete tools for this steering.
Understanding prompting as steering existing capability, not teaching new capability, clarifies exactly when it will and won’t be sufficient — a distinction Article 5 builds on directly.

The Metaphor, Fully Extended

The Archer	What Prompting Changes
Sights that redirect the same archer’s same underlying skill	Prompts that redirect the same model’s same underlying capability
No change to muscles, training, or draw strength	No change to the model’s weights or trained capabilities
Aiming at a different target using unchanged skill	Steering toward a different task using unchanged capability
A quick adjustment rather than months of retraining	A quick prompt change rather than a lengthy fine-tuning run

For Beginners: What to Actually Do

Practice noticing when a prompting change genuinely unlocks better behavior, versus when the model seems to lack the underlying capability entirely, regardless of phrasing.
Learn to think of prompting as steering existing capability, not adding new capability, when deciding whether it’s the right tool for a problem.
Get comfortable exploring this content library’s dedicated prompt engineering series for the concrete techniques behind this steering.

For Practitioners and Leaders: The Deeper Layer

Frame prompting internally as a steering mechanism, setting accurate expectations for what it can and can’t accomplish on its own.
Recognize that a model’s latent capability sets a real ceiling on what prompting alone can achieve, regardless of how well-crafted the prompt is.
Use this distinction to diagnose, early, whether a persistent problem is a prompting issue or a genuine capability gap — the exact question Article 5 addresses directly.

Quick Recap

Prompting steers a model toward using its existing latent capability, without changing its underlying weights at all.
This works because large pretrained models already carry broad, general capability from training.
Prompting is steering, not teaching — a distinction that clarifies its real limits.
This connects directly to the concrete techniques covered in this content library’s prompt engineering series.

Where This Fits in the Series

Article 2 covered what prompting actually changes. Article 3 looks back at how systems were customized before either prompting or fine-tuning, as options, existed at all.