Writing effective prompts
How to write effective prompts
Here are a few guidelines to writing a clear and effective prompt. Read on for more tips and examples.
Be crystal‑clear about “who” and “what”. Start every prompt by defining the model’s role, objective, and scope.
Provide plenty of examples. Show exactly how you want inputs mapped to outputs. For functions, include sample calls and responses under a dedicated “Examples” heading.
Specify output rigorously. Spell out exactly what you need in your results. We support the following values:
Boolean (true or false)
Number
Text
Picklist (a predefined set of options)
Iterate and refine. When you see hallucinations or format drift, add one clarifying sentence to correct it. Test small tweaks and watch how the model’s behavior changes.
General tips for prompts
Tip 1: Always begin by clearly stating the main task or objective you want the model to accomplish.
Objective: Your task is to assess if the provided set of companies are SOC2 compliant.
You are a market‑research analyst with access to up‑to‑date business data. Your goal is to compile comprehensive firmographics on a given company. Respond concisely, cite sources, and flag any data uncertainty.
Tip 2: Keep the main task or objective as short and crisp as possible. Adding unnecessary info would clutter the model’s understanding.
Tip 3: If there are specific conditions or custom business logic you want the model to adhere to, clearly list them under a separate heading labeled “Instructions”. This should solely focus on any specialized knowledge or business requirements you want the model to incorporate.
Instructions:
<List specific business conditions or rules here>
Tip 4: Provide additional examples under Instructions to improve the model’s performance. This part is not necessary but adding additional examples would really enforce the model to follow this pattern and it’s highly recommended for complex use cases.
Instructions:
1. (List specific business conditions or rules here)
Examples:
If the company follows condition_1, then classify it as category_1.
If the company follows condition_2, then classify it as category_2.
Tip 5: If you are requesting multiple outputs from AI Researcher make sure to provide examples and classification instructions for all the labels separately. This should be provided in the “Prompt” area as well as under Label description under “Outputs”.
Advanced tips for prompts
If you have a general idea of where the information might be generally available in a website, try adding that under the Instructions. For example, you want to find all the locations where the company operates in a particular market. You also have a general idea that this information may be present on their website on pages like /about, /locations, /contact.
Adding these mini examples could potentially guide the AI researcher to search for these patterns while browsing. This doesn’t guarantee that the model will look at these pages for all the provided set of companies but it gives a general direction/guideline that the AI Researcher could potentially follow.
Instructions:
The information that you are looking for might be present in pages like:
https://company.com/about
https://company.com/locations
https://company.com/contact
Note: (Sometimes this may hurt the model’s performance if this pattern is not found in majority of the sites, so be sure to experiment with this)
How to format outputs
Provide comprehensive instructions for each output label in the “Output” tab. Clearly define the meaning and purpose of each label under description specifying exactly what the model should identify or classify based on that label.
Ensure consistency by replicating the label instructions in both the “Prompt” and “Description” tabs. This need not be an entire copy paste of all the instructions; just a summarized version should be enough. This repetition strongly reinforces the conditions the model must adhere to during labeling or classification.
When selecting "Picklist" as the value type for your labels, it is strongly recommended to use advanced models such as GPT-4.1.
The new GPT models
AI Researcher now supports the latest GPT-4.1 models. GPT‑4.1 is more literal and obedient to instruction than prior models, making it extremely steerable via well‑specified prompts.
“GPT-4.1 is trained to follow instructions more closely and more literally than its predecessors, which tended to more liberally infer intent from user and system prompts. This also means, however, that GPT-4.1 is highly steerable and responsive to well-specified prompts - if model behavior is different from what you expect, a single sentence firmly and unequivocally clarifying your desired behavior is almost always sufficient to steer the model on course.”
You’ll get more reliable, controllable, and powerful outputs from GPT‑4.1 if you include:
Clear roles/expectations
Structured instructions
Judicious examples
Optional chain‑of‑thought
Last updated