Skip to content
Prompt Corpus
Model behaviour21 August 20262 min read

The prompting advice that stopped working

Prefilling the assistant turn was standard, taught everywhere, and now returns a 400 error. It is the cleanest example of why some technique expires and some does not.

For about two years, one of the most reliable tricks in prompt engineering was prefilling the assistant turn. You supplied the beginning of the model's response — a { to force JSON, or Here are the results: to skip a preamble — and the model continued from there.

It worked well. It was in every guide, including good ones.

On current Claude models it returns a 400 error.

What happened

Anthropic's documentation is direct about it: from Claude 4.6 onward, a prefilled assistant message on the final turn is not supported, and requests containing one fail outright. The stated reason is that instruction following improved enough that most prefill use cases no longer need it.

The documented replacements are worth knowing, because they are better anyway:

  • Forcing a format → structured outputs, or a tool with the schema
  • Killing the preamble → "Respond directly without preamble. Do not start with phrases like 'Here is' or 'Based on'."
  • Continuing an interrupted response → move it into the user turn: "Your previous response was interrupted and ended with [text]. Continue from where you left off."

Adding assistant messages elsewhere in a conversation still works. It is specifically prefilling the last turn.

The thing worth generalising

Prefill worked by exploiting a weakness. Instruction following was imperfect, so putting words in the model's mouth was more reliable than asking. When the weakness closed, the technique died with it.

That is the test to apply to everything you have picked up:

Can you explain why it works in terms of the mechanism?

If yes, it will probably survive. Specificity narrows the distribution. Reasoning has to come before the verdict because tokens condition on prior tokens. Examples transmit form. Fluency is optimised and truth is not. None of those can stop being true without the architecture changing.

If no, you are holding folklore with a shelf life. "Take a deep breath." "I'll tip you $200." "This is important to my career." Some of these measurably helped particular models at particular moments. They are exploits, and exploits expire.

The other one that is quietly expiring

Heavy chain-of-thought scaffolding is going the same way.

"Think step by step, consider all angles, show your working" was written for models that would not reason unless told. Reasoning models do it internally by default, at far greater length than any prompt would elicit. Adding the scaffolding now inflates cost, and Anthropic's guidance explicitly names overthinking as a failure mode you can trigger by over-instructing.

The replacement is the same shape as the prefill replacement: specify the standard, not the procedure. Tell it what a good answer must satisfy and leave the route to it.

If you are migrating prompts to a reasoning model, strip the scaffolding first and re-test. Most of it is doing nothing, and some of it is doing harm.

The full treatment is in chain-of-thought and the Claude playbook.

All notes