Techniques to improve gen AI model output

A retail company wants to customize a large foundation model to consistently follow their internal product-naming conventions and writing style. Their ML team has a modest compute budget and wants to avoid the cost and time of retraining all of the model's parameters. They also want to keep the original base model intact so it can still serve other general tasks. Which customization approach best fits these constraints?

A financial analyst is using a gen AI model to solve multi-step word problems involving budget allocations and percentage calculations. The model often produces the correct setup but arrives at the wrong final number, jumping directly to an answer without showing intermediate steps. Which prompt engineering technique is most likely to improve the accuracy of these calculations?

A financial analyst uses a gen AI model to solve multi-step budget allocation word problems. When she asks the model to give the final answer directly, it frequently makes arithmetic and logic errors. She wants to improve accuracy without collecting labeled examples or tuning the model. Which prompt engineering technique should she apply?

A marketing team uses a gen AI model to generate product descriptions. They have written three different prompt versions and want to systematically determine which one produces the highest-quality descriptions before rolling one out to production. What is the most effective way to evaluate which prompt performs best?

A marketing analyst writes a prompt that includes both instructions ('Summarize the following customer review in one sentence') and the actual review text, all run together in a single paragraph. The model sometimes treats parts of the review as new instructions, producing inconsistent summaries. Without changing models or adding training data, what is the simplest prompt engineering improvement to make the output more reliable?

A product team has deployed a gen AI assistant that drafts customer email responses. Leadership wants a reliable way to measure whether output quality is actually improving as the team iterates on prompts and grounding. Automated metrics alone haven't correlated well with what customers consider a 'good' reply. What is the most effective approach to evaluate and improve output quality in this situation?

A product team at an insurance company has deployed a gen AI assistant that answers policy questions. They frequently adjust prompts to improve responses, but they worry that a change fixing one type of question might silently break answers to other questions. What practice should they adopt to reliably measure whether each prompt change improves overall output quality?

A media company deploys a gen AI tool that summarizes news articles. Early users complain that some summaries are fluent and well-written but occasionally contradict details in the original article. The product team wants to measure this specific problem systematically before their next model update. Which evaluation approach most directly targets the issue they are seeing?

A marketing team uses a general-purpose LLM to generate product taglines. When they simply ask 'Write a tagline for our eco-friendly water bottle,' the outputs are generic and inconsistent in tone. The team wants better results quickly, without collecting a large dataset, training a custom model, or connecting to any external data source. What is the most appropriate technique to improve output quality in this situation?

A financial analyst uses a gen AI model to extract key risk factors from lengthy 10-K filings. With a simple zero-shot instruction, the model produces inconsistent output — sometimes bullet lists, sometimes paragraphs, and occasionally it invents risks not present in the document. The analyst wants to improve consistency and accuracy without the cost and time of collecting a large labeled dataset or standing up new infrastructure. What is the most appropriate FIRST technique to try?

A support operations manager wants a gen AI model to categorize incoming customer tickets into one of five predefined categories using a very specific label format (e.g., 'CAT-3: Billing'). Early tests with a plain instruction produce inconsistent formatting and occasional made-up categories. The manager cannot train or tune a model and needs a quick fix within the prompt itself. Which prompt engineering technique should they apply first?

A marketing analyst wants a foundation model to generate product descriptions that follow the company's very specific style: a punchy one-line headline, exactly three bullet points, and a closing call-to-action. When they simply describe these rules in a plain instruction, the model's output is inconsistent — sometimes it skips the call-to-action or writes four bullets. The analyst does not have a labeled dataset and needs a quick fix without any training. Which prompt engineering technique should they apply first?

A marketing team wants a gen AI model to rewrite product descriptions in their distinctive brand voice. They only have about 15 well-written examples of the desired style, no ML engineers, and need a working solution within a day for a low-volume campaign. Which approach best fits these constraints?

A product team uses a foundation model to convert customer feedback into structured JSON with fields for sentiment, category, and priority. Using a simple instruction prompt, the model returns correct content but with inconsistent key names, varying capitalization, and occasionally missing fields. The team cannot change the model and wants a fast prompt-only fix that enforces a consistent output structure. What should they do?

A financial services company needs its gen AI assistant to consistently produce loan summaries in a highly specific internal format and tone that the base model struggles to replicate, even after extensive prompt engineering with detailed examples in every prompt. They have thousands of high-quality example summaries written by their analysts. Which technique will most reliably bake this behavior into the model so it no longer needs lengthy examples in each prompt?

A healthcare company uses a foundation model to generate patient-facing message drafts. Despite detailed prompts and several examples, the model still fails to consistently adopt the specialized empathetic clinical tone and terminology the organization requires across thousands of daily requests. They have a large curated dataset of high-quality past messages written by their clinical staff. Which approach will most reliably embed this consistent style and vocabulary into the model's default behavior?

A financial services company wants its gen AI assistant to consistently adopt a very specific writing style and reasoning pattern used by its senior analysts. The company has 5,000 high-quality example question-answer pairs written in that exact style. The underlying facts change rarely. Which technique is the BEST fit for teaching the model to reliably produce output in this specialized style?

A financial services company deploys a gen AI assistant that answers employee questions using the firm's internal policy documents through a RAG pipeline. Compliance officers want each answer to include a reference to the specific source document and section it was drawn from, so employees can verify the information themselves. Which capability of a grounded RAG system best supports this requirement?

A financial services firm deploys a gen AI assistant to answer employee questions about internal compliance procedures. Leadership requires that the assistant only respond using content from the company's approved compliance policy documents and refuse to answer when the information is not present in those documents. Which approach best meets this requirement?

A financial services company runs a gen AI assistant that answers employee questions about interest rates, fee schedules, and account policies. These figures change weekly, and any incorrect number could mislead a customer. The team wants the assistant's answers to always reflect the current published figures without retraining the model each week. Which approach best meets this requirement?