Prompt outputs can change sharply after small prompt edits, so product teams need techniques that make responses more consistent and safer.
2
Multi-persona prompting assigns a task to several specialized roles that brainstorm together before producing an answer.
3
According-to prompting grounds an answer in a named source, while emotion prompts add emotional stakes that can improve results on some tasks.
Summary
Dan Cleary presents three prompting methods for improving consistency and accuracy. He starts with the problem: language models are non-deterministic, and small prompt changes can produce very different results. That matters for products because one bad response can reduce user trust, while users expect fast, accurate answers without hallucinations. His first method, multi-persona prompting, has several task-specific roles collaborate on a problem, such as an author, publicist, and target reader working on a book. His second method adds a named source with wording such as "according to Wikipedia" to ground the response. His third method, the emotion prompt, adds emotional stakes to the end of a prompt. Cleary says research found improvements ranging from 8% to 115%, depending on the task. He also points viewers to PromptHub templates for trying these methods.
Small prompt changes can cause large output changes
Cleary says language models are non-deterministic, which makes their behavior hard to predict. He has seen small changes in a prompt produce an outsized effect on the response. This creates a product risk: one time the model goes off the rails can damage a brand or product and reduce trust. Users also approach AI features with high expectations. They want crisp outputs, no hallucinations, speed, and accuracy. Cleary frames prompt engineering as a way to make those outputs better and safer in both everyday ChatGPT use and integrated product features.
Multi-persona prompting makes specialized roles collaborate
Multi-persona prompting calls on several agents to work on one task, with each agent designed for a specific role. Cleary uses writing a book as an example. The model could involve a publicist, an author, and the book's intended audience. These roles brainstorm together, exchange ideas, and collaborate toward a final answer. Cleary says the method is useful when a task is complex or needs additional logic. He personally prefers it for generative tasks, and he values being able to see the collaboration process.
The according-to method adds a specific source to the end of a prompt. Cleary contrasts asking a general question about where starch is digested with asking the same question "according to Wikipedia." He says this increases the chance that the model retrieves information from that particular source. He presents the method as useful when a fine-tuned or general model needs to draw from a consistent source of data. Cleary attributes the research to Johns Hopkins University and says the method can reduce hallucinations by up to 20%.
Cleary's preferred method is the emotion prompt. Research from Microsoft and other universities examined how language models respond to emotional stimuli placed at the end of prompts. His explanation uses a workplace comparison: people may take a product more seriously when a boss says it matters to their career and a major client. The technique adds similar emotional framing to a normal prompt. Cleary says he has seen better and more accurate outputs for tasks such as cover letters and change logs.
Cleary says experiments on emotion prompts found improvements ranging from 8% to 115%, depending on the task. He presents the three methods as quick additions that can be used in ChatGPT or in AI features inside a product. PromptHub provides templates that users can copy, run in its playground, and share with a team. The talk therefore ends with a practical path to testing the methods rather than treating them as fixed prompt rules.
"Adding according to specified Source will increase the chance that the model goes to that specific source to retrieve the information."02:56
Who should watch
You are building an AI feature and need to reduce the damage caused by occasional bad or inconsistent model responses.
You use ChatGPT for writing or research and want simple prompt patterns to test on complex or source-specific tasks.
Your team manages many prompts and wants templates it can run, compare, and share through PromptHub.
