After watching how hundreds of users interact with agents, here are the patterns that consistently produce the best results.
1. Front-Load Context
Bad: "Help me write an email"
Good: "I need to write a follow-up email to a client who attended our product demo yesterday. They seemed interested in the analytics features but had concerns about pricing. Tone should be professional but warm."
Agents work best when they understand the who, what, why upfront.
2. Specify the Output Format
Agents default to prose. If you want structured output, say so:
- "Return this as a markdown table with columns: Feature, Status, Priority"
- "Give me a numbered list, max 5 items"
- "Format this as a JSON object with keys: title, summary, actionItems"
3. Use Step-by-Step Instructions for Complex Tasks
Break complex requests into explicit steps:
"I need you to:
- Read the attached quarterly report
- Extract all KPIs and their YoY changes
- Identify the 3 metrics with the largest negative change
- For each, suggest one concrete action to improve it
- Format everything as a presentation outline"
4. Leverage Agent Specializations
Each agent has a specialty. Match the task to the agent:
- Data Processor for CSV/JSON manipulation, not creative writing
- Creative Writer for narratives, not data analysis
- Software Architect for system design, not code debugging
When in doubt, use a Team — the supervisor handles routing.
5. Iterate, Don't Restart
If the first response isn't quite right, refine in the same conversation:
- "Good, but make the tone more casual"
- "Expand on point #3 with specific examples"
- "Now convert this into bullet points"
The agent retains context and each refinement gets you closer.
6. Use the Library
Upload reference docs (style guides, brand books, data dictionaries) to your Library. Then reference them: "Using the brand guidelines in my library, rewrite this copy."
RAG-powered context beats copy-pasting every time.