Thought-Steering Delimiters

Expert

Inject guiding tokens to steer reasoning directions inside generation streams.

01 The Concept

You can steer a model's generation by pre-filling its response buffer (such as starting the assistant's reply with a specific word or tag).

Thought-Steering Delimiters (like starting responses with `<thought> ` or `{ "status":`) force the model to adopt specific reasoning or formatting routes from its first output token, bypassing conversational introductions.

02 Weak vs. Strong

EX 01JSON response steering

Prompt: Generate a JSON object with keys: name, age. Assistant Response Prefill: {

→ Why it works

Since the model must generate the next token starting from '{', it cannot write conversational introductions, guaranteeing a clean JSON output.

03 Key Points

01Response Pre-filling: Appending the start of the assistant response to the prompt.

02Thought Gating: Starting responses with `<thought>` to force reasoning before the final answer.

03Format Enforcement: Starting responses with `{ ` to guarantee JSON compliance.

04Conversation Bypass: Skipping pleasantries ('Sure, I can help with that!') to save tokens.

05Steering Control: Forcing specific logical starting points to guide subsequent tokens.

04 Model-Specific Notes

Anthropic's API supports response pre-filling explicitly. Use it to force Claude to write in specific XML tags or JSON shapes.

05 For Your Role

It's like putting a pen in someone's hand and drawing the opening bracket of a box, so they are forced to write inside the box.

Next: Token Auditing & Optimization →