Five Proven Levers to Improve LLM Prompts

Making a model's output useful is not about a magic formula but about measurable adjustments. Five techniques, tested on a meeting transcript with reassignments and ambiguities, show where to act: output structure, role, examples, reasoning, and iterative optimization. The goal is not a prettier text but a response that a system can validate and ingest.

Reasoning at Lower Cost and Targeting Ambiguities

Requesting step-by-step reasoning no longer carries the same weight with recent models that already reason internally, unlike the spectacular gains observed in 2022 and 2023. Where it remains crucial is in gray areas, such as the reassignment of a task in a transcript. Without an explicit invitation to reason, a model may retain the first assignment made by Priya and ignore the correction that occurred later. A prompt that requires tracing assignments throughout the exchange and retaining only the final owner specifically targets this ambiguity. To control costs, a variant called Chain of Draft asks for steps of about five words; studies indicate that it can achieve the same accuracy as complete reasoning with only 7.6% of the reasoning tokens, to be used once the value of reasoning is established.

Automating Prompt Optimization Instead of Iterating Instinctively

Rather than refining a prompt blindly, automated and iterative optimization evaluates variants on real test cases and allows a research process to identify useful fixes. Typical fragments show concrete areas to explore: using the final owner when an assignment changes, merging a later integrated task into an existing item without duplication, or marking "unassigned" when no owner is explicitly named. This approach transforms improvement into a measurable, reproducible, and comparable problem.

Composing Examples That Teach Different Patterns

The quality of demonstrations often influences outcomes more than the wording of instructions alone, and combining them is more effective than using either one in isolation. The classic pitfall is gathering three variants of the same pattern, which adds little value. A diverse selection algorithm based on TF-IDF and cosine similarity maximizes the dissimilarity between examples. In a test, naively choosing the first three candidates retained two near-duplicates; conscious selection identified the pair and replaced it with a truly different case. For extracting actions from a meeting, a relevant trio must cover three distinct patterns: a confirmed owner, one explicitly unresolved, and an item merged into a previous one.

Requesting a Precise Format Allows for Automatic Verification

Specifying a measurable output format changes the game. Asking to "list actions" produces readable prose but unusable by a downstream system; in production, the lack of reliable parsing is a failure. By validating a JSON response against a Pydantic schema (ActionItem and ActionItemList) via a dedicated parsing function, the numbered prose fails analysis while the same information, generated according to the schema, passes and materializes into three validated objects. The concrete difference is an output immediately usable by code rather than text that needs to be manually re-entered.

Guiding the Model by a Role Attentive to Reassignments

Framing the system in a specific role activates training behaviors useful to the task. Presenting oneself as a meticulous executive assistant, well-versed in reassignments and prohibited from guessing an owner, leads to more structured outputs than a simple generic command. Conversely, a vague instruction neither signals the need to monitor changes in assignment nor the "unassigned" status when it is explicit, which fosters misinterpretation errors.

The Test Case: A Meeting with Reassignments, Integration, and Unknowns

The evaluation scenario is based on a meeting where three situations complicate the reliable extraction of actions. The review of the mobile layout shifts from Priya to Jake, who commits to finalizing it by Thursday, after Tom indicated the need for review before Friday. The verification of the tablet break point is integrated into this same review, following a previous complaint, and is not isolated as a separate task. Furthermore, the triage of the support queue, which has about 40 tickets to process this week, remains without a designated owner after Priya excluded Tom and Jake and reserved the designation since the rotation. The processing of the billing migration is on hold until Tom reads the changelog of a recent security fix, which he says he plans to do the next morning. These variations and unknowns form the testing ground for the previous strategies.

Why These Techniques Apply to Optimization, Not Design

The distinction between engineering and prompt optimization is not theoretical. Most teams already have a useful prompt and seek concrete gains without touching the model. The five levers described—output structure, role, example selection, targeted reasoning, and iterative optimization—fit precisely within this framework: they refine an existing prompt according to verifiable criteria and tangible measures.