Match Prompt Structure to PM Task Type and Decompose Hybrid Pipelines

Match Prompt Structure to PM Task Type and Decompose Hybrid Pipelines

The previous lesson covered selecting examples. This lesson tackles a bigger question: what the overall prompt structure should look like, depending on the nature of the task.

Running Example: The Product Manager of FreelanceFlow needs AI to score PM candidate resumes against a hiring rubric, and to read transcripts from 10 customer interviews to write an Executive Product Opportunity Brief for Leadership.

1. Enforce Hard Rules and Strict Formats for Precision Tasks

There is no universal prompt template for all PM tasks. When executing Precision tasks (resume scoring, financial metric extraction, classification), prompt architecture must follow strict engineering standards:

• Eliminate Title Personas: Do not open with "You are a seasoned HR director...", as this persona promotes leniency and subjective storytelling.
• Define Unambiguous Rubrics: Provide clear binary or tiered thresholds for acceptance/rejection.
• Constrain Output Enums: Specify a closed set of allowed values (preventing the model from inventing creative labels).
• Suppress Reasoning Output Unless Auditing: If downstream code parses the output, explicitly forbid conversational commentary outside the required schema.

Example: Scoring PM resumes for FreelanceFlow:

## Task
Evaluate the Product Manager resume against the 3 criteria below. Score each on a 1-5 scale (1: No evidence, 5: Exceptional with verified metrics).
 
## Rubric
1. Product Analytics: Proven experience with tracking plans and funnel analysis (Mandatory).
2. AI/LLM Execution: Track record of authoring PRDs or deploying production AI features.
3. Stakeholder Communication: Clear evidence of cross-functional alignment.
 
## Output format
Return valid JSON matching this schema:
{
  "scores": { "analytics": number, "ai_execution": number, "communication": number },
  "total_score": number,
  "passed_gate": boolean
}

2. Encode Tone and Behavioral Boundaries for Generative Tasks

In contrast to Precision tasks, Generative jobs (drafting LinkedIn launch posts, writing incident announcements, answering stakeholder inquiries) do not have a single objective ground truth. Quality is dictated by Tone & Voice, Target Audience Alignment, and Brand Safety Guardrails.

Prompt architecture for Generative tasks should focus on:

• Functional Persona: Clarify the spokesperson role and strategic perspective.
• Target Audience: State reader expectations, knowledge depth, and core pain points.
• Negative Constraints: Explicitly forbid hollow buzzwords or unsafe promises (e.g., ban terms like "game-changing", "revolutionary", "seamless").

Exercise 2.2: Classify the following 3 PM tasks into Precision or Generative, then draft the corresponding prompt structure:

1. Extracting all unresolved questions from a 5-page Sprint Review transcript.
2. Drafting a Push Notification for a "Budget Limit Exceeded Warning" targeted at freelancers onboarded within the last 7 days.
3. Categorizing 100 NPS survey responses into 4 buckets: [UI/UX], [Pricing], [Bug], [Feature Request].

3. Decompose Hybrid Tasks into a 2-Step Pipeline Instead of a Monolithic Prompt

In real-world product workflows, many responsibilities are hybrid: they demand rigorous data extraction combined with persuasive narrative synthesis. A classic example is: "Read transcripts from 10 customer interviews and write an Executive Product Opportunity Brief for Leadership."

Shoving both goals into a single prompt forces an attention conflict: the model struggles between strictly adhering to verbatim data points (Precision) while simultaneously generalizing and writing engaging prose (Generative). The outcome is often mediocre - inaccurate metrics wrapped in dry, robotic writing.

Engineering Solution: Always decompose hybrid tasks into an independent 2-step pipeline:

Step 1 (Precision Prompt): Extract all pain points, recurrence counts, and verbatim quotes into a clean structured table. Step 2 (Generative Prompt): Ingest the Step 1 data table to draft the Executive Summary, leveraging a functional leadership persona and strategic framing.