Practical Prompt Patterns for PM Workflows

I have been using AI models daily for PM tasks since mid-2024. In that time, I have distilled my approach into five prompt patterns that consistently produce useful output. These work with Claude 3.5 Sonnet, GPT-4o, and most capable models.

Pattern 1: The Structured Analyst

For risk analysis, I give the model a specific analytical role. "You are a senior program manager reviewing this project plan. Identify the top five risks, rate each on likelihood and impact (high/medium/low), and propose one mitigation action per risk. Format as a table." This consistently produces better output than simply asking "what are the risks."

Pattern 2: The Devil's Advocate

Before any major decision, I ask the model to argue against my plan. "Here is my proposed approach to staffing this program. Argue against this approach. What am I missing? What assumptions am I making that could be wrong?" The model is remarkably good at finding blind spots when you explicitly ask it to.

Pattern 3: The Translator

I frequently need to translate technical concepts for non-technical stakeholders. "Rewrite this technical architecture decision for a VP of Operations who has no engineering background. Preserve the key trade-offs but remove all jargon. Keep it under 150 words." This saves me 20 minutes per stakeholder communication.

Pattern 4: The Template Generator

For repeatable artifacts, I provide two to three examples and ask for the next one. Sprint summaries, release notes, and client updates all follow consistent formats. Few-shot prompting with real examples maintains my voice and structure.

Pattern 5: The Edge Case Finder

When reviewing requirements, I ask: "Read this feature specification. List every edge case, boundary condition, and ambiguous requirement you can identify. For each, suggest a clarifying question to ask the product owner." This catches things that slip through human review.

The Meta-Pattern

Every effective prompt has three components: a specific role for the model, clear context about the task, and explicit constraints on the output. Master these three, and you can prompt your way through almost any PM workflow. The models are capable. The quality of your output depends on the quality of your instructions.