USE FURNACE / NO GPU REQUIRED
Check your prompt prefix.
See whether a prompt edit shortens the stable prefix shared across requests. Start with an example, then compare your own rendered prompts.
Private by design. Comparison runs in your browser. Prompt text is not uploaded, saved by Furnace, or sent to a model. Clear prompts when finished; browser extensions may have their own access.
Paste rendered requests as JSONL: one object per line with a messages array of text role / content fields. Use 2–200 distinct final user questions, matched in the same order in both snapshots. Limit: 2 MiB each. Examples above use synthetic prompts.