AI Product Management System
Role
Design Engineer
Duration
Continuous project
Responsibilities
Architecture, Agents, Prompt Engineering
Focus
AI-First & Product Management
CONTEXT
Transforming product management through practical AI
During an internal restructuring process at Flexpro Sistemas, an initiative emerged to incorporate AI practically into workflows. Product management was identified as a critical area where much of the time was consumed by operational and unstructured tasks.
THE PROBLEM
PM work showed clear patterns of inefficiency: undocumented and hard-to-track decisions, poorly structured tasks reaching the development team, and high rework due to a lack of contextual clarity.
Ineficiência
Discovery based on opinion rather than a structured process.
Rework
Constant need to re-explain context to developers.
OBJECTIVE AND APPROACH
The goal was to build a system that reduced operational effort, structured decisions, and maintained continuous product memory. Instead of using AI as a one-off tool, I designed a system where context is persistent and knowledge evolves with use.
ARCHITECTURE DECISION
The primary decision was to build a multi-agent system. This allowed for reduced token consumption, increased accuracy, and context activation only when necessary. The architecture was divided into four fundamental pillars:
Context
Structured and persistent product knowledge.
Agents
Specialized behaviors for each type of task.
Knowledge
Live and institutional memory of the system.
Tasks
Structured, documented, and reusable output.
AGENT SYSTEM
Each agent operates in a specific PM workflow: Discovery, Decision, Task-detail, Innovation, Improvements, and Error-mapping. This transforms the AI from a generic assistant into a context-specialized operator.
AI IN PRACTICE
Claude was used as active infrastructure for both development and execution. This included prompt refinement directly within the workflow and the iterative structuring of the file system to support the agents.
RESULTS AND IMPACT
Speed
Estimated savings of ~2.5h per week in task writing.
Quality
Drastic reduction in development team rework.
Memory
Total reduction in reliance on tacit knowledge.
LEARNINGS
AI performs better when structured as a system, not as an isolated tool. Context is the primary asset in LLM workflows, and breaking down problems into agents drastically increases the predictability of responses.