# The AI PM Starter Kit
Product Manager Coach · September 2026

Use approved workspaces and replace example content with your own evidence.

## 1. Context brief
- Decision or task:
- Audience and user goal:
- Evidence with source IDs:
- Constraints:
- Known facts:
- Assumptions:
- Required output:
- What good looks like:
Example: New users cannot identify their next onboarding step. We have five interview notes (I1–I5). Explore two flows. Use sample data. Do not claim all users have the problem.

## 2. Interview synthesis
For every finding record: source ID; exact excerpt; observation; interpretation; conflicting evidence; uncertainty; follow-up question.
Example: I2, “I thought I was finished.” Observation: participant stopped before the last task. Interpretation: completion state may be unclear. Next: test a clearer next-action prompt.
Prompt: Use only the supplied interviews. Link every finding to a source excerpt. Separate observations from interpretations. Identify conflicting evidence and missing context. Do not invent quotes or imply population prevalence.

## 3. Sketch annotations
- Number each screen.
- Label actions and navigation arrows.
- Mark required content and optional layout ideas.
- Include important empty, loading, error and success states.
- Name the audience, task and test question.
Example annotation: Screen 2 → button “Continue setup” → Screen 3. Progress is informational; user can return to previous steps. Use sample data.
Prompt: Describe the attached sketch before building. Identify ambiguity. Separate supplied requirements from assumptions. Ask about missing details that affect the flow.

## 4. Decision memo
- Decision to make:
- Goal and why now:
- Evidence and source links:
- Options (including no change):
- Tradeoffs and uncertainties:
- Recommendation:
- Next test and owner:
- Evidence that would change the decision:
Example: Test a clearer progress view before building a guided tour. Existing interviews suggest confusion, but do not establish its prevalence. Reconsider after formative testing and a review of funnel data.

## 5. Workflow quality check
- Source permissions respected?
- Personal data minimized?
- Stable record IDs and safe retries?
- Quotes and claims traceable?
- Required output fields validated?
- Ambiguous records sent for review?
- Human approval before consequential writes?
- Failures visible to an owner?
- Total preparation plus review time measured?
- Evaluation examples checked after prompt/model changes?
Example: Reject a feedback summary whose supporting quote cannot be found in the original record. Log the reason and route it to review.