After a well-meaning intern “improves” the canal-inn prompt, reservation answers get worse. What should the team do instead of starting a fine-tune?
Select an answer to reveal the explanation.
Short Explanation
Think of an intern improving the prompt until reservation answers get worse. Roll back to the last good version. A fine-tune is not the undo button.
Full Explanation
When a well-meaning intern “improves” the canal-inn prompt and reservation answers get worse, the first operational move is to restore the last known-good prompt version rather than escalate to a fine-tune. Prompt versioning exists so drift can be undone quickly: roll back the text, then re-evaluate quality on a fixed set of reservation scenarios. That works because the regression is an operations and prompt-control problem, not proof that the foundation model must be customized. Treating prompt drift as a training job wastes time and may bake the bad instructions into customized weights. Deleting every version removes the undo path and turns a recoverable mistake into permanent loss of the working template. Raising temperature may change fluency but does not restore the previous instruction set that produced better answers. Exam caveat: rollback fixes the prompt text; if the intern also changed tools, RAG sources, or guardrail configs, those must be rolled back or audited as well. Operational check: after rollback, run a fixed reservation regression set and compare answer quality to the pre-change baseline before allowing further prompt edits. Require a short change note on every prompt publish so the inn knows what to reverse if quality drops again.