[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"glossary-prompt-drift::en":3,"gloss-cluster-prompt-drift::en":20,"gloss-next-prompt-drift::en":9},{"slug":4,"category":5,"name":6,"definition":7,"meta_desc":8,"faq":9,"schema_markup":9,"related":10},"prompt-drift","prompt-eng","Prompt Drift","Prompt drift refers to the gradual change — usually degradation, though sometimes just unexpected behavior change — in a prompt's real-world output quality or consistency over time, even when the prompt text itself has not been edited, most commonly caused by an underlying change outside the prompt: the model provider silently updating or deprecating the model version behind an API endpoint, shifts in the real-world distribution of inputs the prompt now receives (a support-ticket classification prompt tuned on last year's common ticket types facing this year's different mix of issues), or the accumulation of edge cases and unusual inputs that a prompt's original design never anticipated. This is a distinct failure mode from a deliberate prompt-versioning change (which is a tracked, intentional edit) — prompt drift is often invisible until accuracy metrics or user complaints reveal something has degraded, making it a genuine production reliability risk for any team not actively monitoring AI feature quality over time. Model-provider-caused drift is a particularly important and under-appreciated version of this problem: providers periodically update models behind the same API model identifier (or deprecate older versions on a timeline), and a prompt carefully tuned and evaluated against one model version can behave meaningfully differently — sometimes better, sometimes worse, sometimes just different in a way that breaks a downstream parser expecting a specific format quirk — once the underlying model changes, even though not a single character of the application's prompt code was touched. Defending against prompt drift requires ongoing, not one-time, evaluation practice: re-running the prompt-testing evaluation set on a recurring schedule (not just at initial launch), pinning to specific model versions rather than a \"latest\" alias where reproducibility matters and consciously re-evaluating before adopting a new model version, monitoring real production output quality signals over time (user feedback\u002Fcomplaint rates, downstream parsing failure rates, sampled human review), and treating a model-version upgrade with the same evaluation rigor as a deliberate prompt-version change, since from a quality-assurance perspective the two are functionally similar risks. Concrete worked example: an AI content-moderation feature's classification accuracy, stable for months on the same unedited prompt, unexpectedly drops after the underlying model provider deprecates the pinned model version and auto-migrates traffic to a newer default — the new model interprets one ambiguous category boundary slightly differently than the old one did, and because the prompt hadn't changed, the team initially struggles to explain the accuracy dip until they check their model-version logs and correlate the drop precisely with the migration date. The fix — re-running their evaluation set against the new model version and adjusting a few category-boundary examples in their few-shot prompt to realign — restores accuracy, but the underlying lesson drives a process change: pin model versions explicitly and re-evaluate before any version migration, rather than accepting silent \"latest\" auto-upgrades.","Prompt drift is a prompt's gradual real-world performance decline over time, often caused by silent model updates rather than the prompt itself changing.",null,[11,14,17],{"slug":12,"name":13},"in-context-learning","In-Context Learning",{"slug":15,"name":16},"prompt-testing","Prompt Testing",{"slug":18,"name":19},"prompt-versioning","Prompt Versioning",[21,25,28,31,35,38,41,44,47,50,53,56],{"slug":22,"category":5,"name":23,"updated_at":24},"analogical-prompting","Analogical Prompting","2026-08-24T02:46:37+00:00",{"slug":26,"category":5,"name":27,"updated_at":24},"automatic-prompt-optimization","Automatic Prompt Optimization",{"slug":29,"category":5,"name":30,"updated_at":24},"chain-of-density","Chain of Density (CoD)",{"slug":32,"category":5,"name":33,"updated_at":34},"chain-of-thought-prompting","Chain-of-Thought Prompting","2026-08-24T02:46:36+00:00",{"slug":36,"category":5,"name":37,"updated_at":24},"chain-of-verification","Chain-of-Verification",{"slug":39,"category":5,"name":40,"updated_at":34},"chunking","Chunking",{"slug":42,"category":5,"name":43,"updated_at":34},"constrained-decoding","Constrained Decoding",{"slug":45,"category":5,"name":46,"updated_at":34},"context-stuffing","Context Stuffing",{"slug":48,"category":5,"name":49,"updated_at":34},"delimiter","Delimiter",{"slug":51,"category":5,"name":52,"updated_at":24},"directional-stimulus-prompting","Directional Stimulus Prompting",{"slug":54,"category":5,"name":55,"updated_at":24},"emotion-prompting","Emotion Prompting",{"slug":57,"category":5,"name":58,"updated_at":34},"few-shot-prompting","Few-Shot Prompting"]