Skip to content
Home › Prompt Engineering & Structured Output

Prefill Migration ⚠ verify

Prompt Engineering & Structured Output

Definition

Prefilling the response (a partial assistant message to continue from) on the last assistant turn is no longer supported on newer models; requests that do so return a 400 error, so older prefill patterns must be migrated.

Key points

  • Starting with Claude 4.6 models (and Claude Mythos Preview), a prefilled assistant message on the LAST turn returns a 400 error. ⚠ verify (2026-07-08)
  • Earlier models still support prefills; adding assistant messages elsewhere in the conversation is unaffected.
  • Migration paths by scenario:
    • Controlling output format (JSON/YAML/classification): use Structured Outputs; or ask the model to conform (newer models match complex schemas when told, esp. with retries); for classification use a tool with an enum of valid labels.
    • Eliminating preambles: direct instruction ("Respond directly without preamble. Do not start with 'Here is...', 'Based on...'"); or output within XML tags / structured outputs / tool calling; strip stray preambles in post-processing.
    • Avoiding bad refusals: clear prompting in the user message now suffices.
    • Continuations: move continuation into the user message ("Your previous response was interrupted and ended with [previous_response]. Continue from where you left off.").
    • Context hydration / role consistency: inject former prefilled reminders into the user turn, or hydrate via tools / during compaction.

Why it matters for the exam

  • "Last-turn prefill → 400 on 4.6+" is a crisp, testable version fact.
  • Recognizing Structured Outputs / tool enum as the replacement for format-forcing prefills is the tested reasoning.

Common gotchas

  • The 400 applies only to the last assistant turn — earlier-turn assistant messages are fine.
  • "Keep using prefill to force JSON" is wrong on 4.6+; migrate to Structured Outputs.

See also

Sources

Referenced by