A Prompt on Burnout for AI "Reasoning" Models
March 17, 2025
Some analysts believe that newer LLM AI models need a different prompting structure. As Jordan Gibbs explains in his post, Prompt Engineering Has Changed Forever -- Learn How to Adapt, many newer models are "reasoning" models -- they think before responding rather than just beginning with a word and then allowing the salad to come out essentially based on a "what might we expect to find next if we looked at millions of phrases" query for each succeeding word. These models include ChatGPT's 01, 01-mini, and 03-mini, to name a few. As comfortable as I have been with using my "define a role and then provide a task" approach, Gibbs suggests that one should use the following structure when dealing with reasoning models: state the goal, the output structure you want, things the model should avoid, then the context [that is, the data you want to have the model review, including websites to review and documents to evaluate]. End with a request that the model ask any clarifying questions it needs. So according to Gibbs, the structure should be:
[Your goal statement + any clarifying information about the goal]
— — — — — — — — — — — — — — — — —
[Your output structure statement]
— — — — — — — — — — — — — — — — —
[Your statement about things to avoid (you can delete this section if there are none)]
— — — — — — — — — — — — — — — — —
[Dump all of your context here]
— — — — — — — — — — — — — — — — —
What might this look like for a nonprofiteer? How about a task to gather best practices for preventing nonprofit employee burnout? Try this one, using ChatGPT 03-mini [if you don't know how to choose a model on ChatGPT, look here]:
Provide a force-ranked list of recommendations about how to manage nonprofit employees to avoid burnout. To the extent relevant, focus on the particular challenges faced by nonprofit employees.
Provide results in a table listing the following columns: Recommendation, Reason, and Source, where "source" is the name of the source and a hyperlink to that source
Avoid any recommendations that would require me to raise salaries or improve financial benefits. Act on the assumption that I work at a nonprofit that does not have the financial resources to make substantial changes in this regard.
Perform substantial online research of sources on the internet that provide HR recommendations to nonprofit leaders. Supplement that with additional research on reputable medical websites like Mayo Clinic, Harvard Health, and WebMD, as well as reputable sources for psychological advice like PsychCentral, Psychology Today, and Scientific American Mind & Brain.
I'm playing around with this approach. I will report back on whether it is useful or not. I will say that the output for the prompt above was helpful for me.
If you think this AI journey might be useful to anybody else, please forward this newsletter to them and suggest that they subscribe using this Link. By multiplying the readers of this newsletter, you again help build our common resilience and bring new minds to bear on these issues. Let’s grow this community together!