Assistant Message — AI Glossary

Reference details

Related terms

• User Message
• System Message
• Tool Definition

How it works

1. The chat format is a list of role-tagged messages that the serving layer flattens into one token sequence using the model's chat template.
2. An assistant message is the slot the model's own output occupies, so placing text there asserts that the model previously said it.
3. Because the model continues from the end of the sequence, a partial assistant message at the tail is completed rather than answered — this is prefill, and it is how output format is forced.
4. Tool calls arrive inside an assistant message as structured arguments; the caller executes them and returns results in a separate tool-role message the model reads on its next turn.
5. Nothing verifies the history, so a caller can write an assistant turn the model never produced and the model will treat it as its own prior commitment.

Example: Prefilling to force a format

Asked for JSON, a model often opens with 'Here is the JSON you requested:' and the caller has to strip it. Ending the request with an assistant message containing just the opening brace makes the next token a continuation of that brace, so the response begins inside the object. No instruction achieved it; the position in the sequence did. The same mechanism is what makes a fabricated assistant turn effective — and why one should never be treated as a record of what happened.

Commonly confused with

The model's turn in the transcript — but nothing stops you writing one yourself. Placing text in an assistant message makes the model continue from it as though it had said it, which is a steering technique rather than a record of what happened. So an assistant message in a request is a claim about the conversation, not evidence of it, and traces should distinguish the two.

When to use it

Reach for it when:

• Forcing an output format by prefilling the opening tokens rather than asking for it in prose.
• Supplying few-shot examples in chat format, where each exemplar needs a user and an assistant turn.
• Reasoning about a tool-calling loop, where the call lives in an assistant message and the result does not.

Reach for something else when:

• As evidence of what a model actually said; a transcript can be authored.
• For instructions that should carry authority — those belong in the system message.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Assistant Message — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path