# No customer fine-tuning of Claude

You cannot retrain Claude's weights on your data; you adapt behavior through prompts, context, tools, and RAG.

## Why the CCAR-P exam tests this

It is a heavily-tested misconception — candidates wrongly assume they fine-tune Claude like an open-weight model.

## Definition

As a customer you do **not train or fine-tune Claude's weights** on your own dataset. Instead you steer the model at **inference time** using everything that goes into the request: the **system prompt**, few-shot **examples**, retrieved **context** (RAG), **tool** definitions, and output constraints.

This matters because the instinct "the model doesn't know my domain, so I'll fine-tune it" is the wrong lever here. The correct fixes are supplying the domain knowledge in context (via RAG or documents), giving clearer instructions and examples, and wiring tools so the model can fetch or act on live data. These are cheaper, faster to iterate, and auditable.

Behavior you get from good prompting and retrieval is often mistaken for "training" — but nothing about the model's parameters changes. Each request is independent; there is no persisted learning from your prior calls unless you re-supply that context yourself.

## Exam trap

If an answer says "fine-tune Claude on the company's tickets," it is wrong — the intended fix is RAG, better prompting, examples, or tools, not weight training.

## Commonly confused with

- **context-window-management** — Adding knowledge means putting it in the context window at inference, not baking it into weights.
- **prompt-caching** — Caching a stable system prompt speeds reuse of steering context but is still not fine-tuning.

## Related

- model-selection-opus-sonnet-haiku
- context-window-management
- structured-output

---

Source: Claude Cert Prep — an independent, unofficial CCAR-P study resource (domain P2). Not affiliated with Anthropic.