Claude Code workspace for product management

Result delivered

One workspace ran all four of my mandates as Head of Product:

• Logistics. I launched and scaled the logistics program, a vertically integrated logistics business, growing merchants nearly 14× and monthly orders 5× within four months of its pilot and reaching about 20% of weekly marketplace gross merchandise value. I lifted on-time delivery 18+ points in two weeks after volume grew about 4×, and learned Colombia’s geography to improve delivery times. See the control tower and the enrollment plan.
• Customer experience. I deployed the in-app AI support agent and gated its rollout on evaluations. It held at 10% until six fixes lifted correct escalations from 65% to 90%, then shipped to all customers at 95% and cut outsourced support volume 60%. See the agent evaluation.
• Product management. I managed the customer support product manager and a head of operations with a team of two and one open role, with a commercial lead, a program manager, four engineers, and data science on a dotted line. I built the logistics operations team around AI-automated checks, wrote the team’s norms and ways of working, and fed the engineering roadmap in Linear from one reconciled plan. See the roadmap coordination.
• Executive team deliverables. I set the merchant fee and settlement models with Finance, and settled their principal-or-agent tax treatment under Colombian law with the general counsel and outside counsel, with no prior compliance or regulatory background. Every leadership document passes written checks before it publishes.

Background

I held four mandates at once, each with its own tools, meetings, and documents. My first workspace grew to 240,789 lines of instructions and checks, and the same mistakes still repeated.

Action taken

I cut the workspace to 39,851 lines and rebuilt it around one local folder in git. Claude reads it at the start of each session and writes every correction back, so a mistake fixed once stays fixed. It has:

• One folder per workstream, with instructions kept apart from current state.
• 10 skills for routines done more than twice, such as weekly planning and Linear status updates.
• A weekly maintenance run that adds missed corrections and removes outdated instructions.
• An approval bar, under which Claude acts on anything recoverable and asks before any send.

How I used AI

I ran it in Claude Code in Visual Studio Code, on Opus at the extra-high effort setting. Connectors gave each session the team’s metrics, delivery tracker, documents, meeting notes, and messages, shown as icons in the flowchart above.

Decisions and trade-offs

I first tried to match my pace of work and decisions with rigid rules. Designing for resilience and variability mattered more. Once I designed for them, the workspace shrank to about a sixth of its size, the checks that stayed stopped failing the same way, and a mistake fixed once now stays fixed. Three decisions made that work:

• I kept only the checks that proved they worked. The document checks had zero repeat failures in 22 days, so they stayed and the rest retired.
• I pay for any new safeguard by cutting something elsewhere, so the workspace stays small.
• I keep Claude’s built-in memory empty, so every durable fact sits in a file git tracks.