DeepMind researchers detail Optimization by PROmpting to improve LLM performance by using “meta-prompts” like “take a deep breath”, which helped Google's PaLM 2
DeepMind used AI models to optimize their own prompts, with surprising results.
Context & Ripple Effects
Google had positioned PaLM 2 around multilingual, reasoning and coding improvements in its initial PaLM 2 rollout. This research adds a different performance lever: optimizing the instructions around a model rather than changing the model itself.
Related coverage later found that emotionally framed prompts could also improve model responses, reinforcing that prompt context can materially affect output quality. The significance is that prompt design becomes a repeatable technical layer, not merely ad hoc user phrasing.
First-order effects
- DeepMind and Google gain a documented method for using an LLM to search for prompt formulations that improve PaLM 2 task performance.
- Teams deploying LLMs can treat the prompt as an optimizable component of the system, including instruction styles that may look superficial but alter results.
Second-order effects
- Model providers and application builders face pressure to benchmark prompt-optimization methods alongside model upgrades, because better instructions can change perceived capability without a new base model.
- Prompt tooling and evaluation become more valuable to customers: a model’s practical performance increasingly depends on the task framing and test set used around it.
Third-order effects
- If such gains generalize reliably, differentiation will shift partly from raw model quality toward context-engineering workflows that encode task knowledge, evaluate outputs and continuously revise instructions.
- The pattern also makes reproducibility more important: capability claims may be harder to compare when two users of the same model operate it with materially different optimized prompts.
The trend: This is one data point in the shift from treating prompts as user inputs to treating context engineering as an operational layer of LLM performance.