How Claude Thinks Silently

This shows how a model solves problems using its internal workspace — thinking first and reasoning before it speaks — so not everything appears in the output.

How it works

1. 1Read and understand the prompt.
2. 2Think silently in the internal workspace (concepts, intermediate steps).
3. 3Do the reasoning or calculations internally.
4. 4Decide on tools or actions if needed.
5. 5Speak only the final answer.

Key points

• Example: for '2 + 5 × 3', intermediate steps happen silently; only the final answer is written.
• The internal workspace holds the reasoning; only the result is output.
• One internal concept can be reused for many tasks.
• Thinking is invisible — understanding it changes how we monitor models.

Why it matters

Because much reasoning is silent, we can't judge a model only by its final words. Studying hidden thinking is essential for safety and debugging.

Frequently asked questions

Why don't I see all the steps?

Much of the reasoning happens internally; the model typically writes only the final answer unless asked to show work.

Can we see the hidden steps?

Interpretability research aims to read internal activity to reveal them — that's the point of this series.