by.waclaw.online / pm-agent

The Product Manager Agent

An agentic assistant that turns vague asks into testable requirements — and a documentation librarian that keeps a living catalog of what your internal app actually is. Designed for enterprise internal apps in the CPG industry, grounded in Claude Code, Jira, and docs-as-code on GitHub.

The one-paragraph idea

Two failures bleed time on every internal-app team. First, requirements arrive vague — "we need better promo visibility" — and the cost of that vagueness is paid later, as rework, missed edge cases, and three more meetings to pin down what "done" meant. Second, the documentation that would prevent that vagueness — the data model, the screens, the batch jobs, the approval rules — rots, because keeping it current is everybody's job and therefore nobody's. This guide designs an agentic system that attacks both at once: an Analyst agent that grills a fuzzy idea into a crisp Jira story using what the system already knows, and a Librarian agent that maintains that knowledge as a living, continuously-updated catalog — a wiki in the Karpathy sense, never a frozen spec.

It is built on tools you already have. Jira is the system of record for user stories. GitHub holds the documentation as code — Markdown, Mermaid, and schema files that change through pull requests — alongside the agent's own configuration. The reasoning layer is Claude Code / the Claude Agent SDK: subagents, skills like grill-me, and MCP connections to Jira and GitHub. We design the whole thing around one concrete running example: PromoDesk, an internal Trade Promotion Management app for a CPG company.

Read the version that fits you

This guide ships in two layers. They tell the same story at different altitudes — read one, or both.

For leaders the case

The problem in business terms, what changes when this exists, the ROI and the metrics to watch, the risks, and an adoption path. Code-free, one read.

Read the leadership case →

For builders the build

The full design: agent topology, the canonical entity catalog, the elicitation skills, the Jira + GitHub wiring, and a Claude Code reference implementation. Start at chapter 1 below.

Start the build →

The build, in eight chapters

  1. The Problem: Vague Requirements and Documentation Rot concept
  2. The Architecture: Orchestrator, Analyst, Librarian concept
  3. The Entity Catalog: A Living Map of the System concept
  4. The Librarian: Keeping the Catalog Alive build
  5. Elicitation: The Analyst and the grill-me Skill build
  6. The Backend: Jira for Stories, GitHub for Docs-as-Code build
  7. Reference Implementation in Claude Code build
  8. Adoption, Metrics, and Guardrails concept
A note on the running example. Every chapter uses PromoDesk, an internal Trade Promotion Management application for a mid-size CPG company — the kind of app where key-account managers plan retailer promotions, the system accrues trade spend into a fund, and retailer deduction claims have to be reconciled against what was promised. It is chosen because it exercises every part of the catalog: real screens, a real SQL schema, nightly batch jobs, threshold-based approvals, and ERP/EDI integrations. Map it onto your own app as you read.
A note on recency. Agent platforms move fast. The names of Claude Code primitives, the exact shape of the Jira and GitHub MCP servers, and a given Atlassian endpoint will drift from what is written here. Treat the shape as stable — two specialist agents, a docs-as-code catalog, story-to-doc traceability, skills like grill-me — and check current tool details with the provider before you ship.