In our previous post , we explored delegation —understanding the problem, selecting the right LLM, and deciding which tasks are human-led versus AI-led. Now, we shift focus to description : how to effectively communicate the tasks you've delegated to an LLM, ensuring clarity, precision, and actionable results.
Consider a scenario: a colleague asks, “Can you help me generate report using AI?” At first glance, it seems simple. But without a clear description, the AI might produce irrelevant or low-quality outputs.
To avoid this, we break description into three distinct layers:
Purpose: Clearly articulate what you want the AI to produce.
Key Elements: Format, audience, style & tone, content details.
Example: "I am a technical lead that want to generate a market analysis of observability and monitoring tools for software development services to present to a technical audience. It should include a comparison table of features."
Purpose: Instruct the AI on how to tackle the task.
Key Elements: Methodology, reasoning style, interaction style.
Example: "Build the report around the Weighted Scoring Model framework, evaluating tools based on criteria like ease of use, integration capabilities, scalability, and cost-effectiveness. Start with an executive summary, followed by detailed sections for each tool, and conclude with a recommendation."
Purpose: Set expectations for how the AI should behave while performing the task.
Key Elements: Level of detail, perspective, engagement style.
Example: "Provide a detailed, supportive explanation with a focus on inclusivity and clarity. when I ask questions be challenging against my assumptions to help me think more deeply about the topic."
To delegate effectively, ensure all three layers are addressed:
Mastering description is the next step after delegation. By layering product, process, and performance instructions, and following the actionable takeaways, you can collaborate with AI more deliberately, reduce errors, and produce outputs that truly meet your goals.
Sources and further reading: