Direct Wines — AI across a 70+ person e-commerce org
An AI product-development workflow and coding guardrails embedded in the team's existing stack.
Fractional product lead · Direct Wines · concluded March 2026
- PROBLEM
- A 70+ person engineering organisation needed a practical AI workflow embedded in its existing delivery stack.
- BUILT
- An AI product-development workflow plus repository rules, custom agents and skills, pre-commit checks, and CI controls.
- OUTCOME
- 70+ person engineering org · < 1 hour benchmark implementation · 2 devs / 10 days original delivery benchmark
- STACK
- Claude API · Claude Code · GitHub Actions · Cursor · VS Code · GitHub Copilot
- STATUS
- handed over — it runs without me
Why it exists
Direct Wines had a 70+ person engineering organisation shipping across multiple platforms. The delivery workflow needed a practical way to turn business context into review-ready product requirements and engineering-ready implementation prompts, with guardrails that worked inside the existing stack.
What I built
I built and piloted the workflow with the delivery team, integrating Jira, Confluence, GitHub, and Claude/OpenAI APIs. Three parts turned the source context into something the team could use:
- Context becomes a buildable brief. Business requirements, tickets, documentation, and repository context become review-ready PRDs, user stories, and implementation prompts.
- Rules the agents obey. Repository rules, custom agents, and skills carry the codebase’s architecture and high-risk boundaries into the coding tools.
- Checks in the delivery workflow. Pre-commit checks and CI controls keep the guardrails active through implementation and review.
Every pull request gets a review like this:
🤖 AI Code Review — Quality 8/10
🔴 CRITICAL checkout.ts — cart call missing cookie forwarding
🟡 WARNING ProductCard.tsx — UI component must not fetch data
🔵 SUGGESTION PriceTag.tsx — add a co-located testThe result
In a prepared benchmark, the workflow produced a test-passing implementation in under one hour against the same requirements and acceptance criteria as work originally delivered by two developers over 10 working days. I also rolled the coding guardrails out across the 70+ person engineering organisation.
Alongside the AI work, I led a six-person team that consolidated the cloned storefronts into a multi-tenant commerce platform. That work and its outcome metrics are covered in the separate platform case study. The engagement concluded in March 2026 and the system was handed to the team. The full AI breakdown is on fractalocean.ai.
“He also brought AI into the team in a way that actually stuck. From a Claude-powered PR review bot to AI-driven rule systems embedded in the IDE, these weren’t experiments, they’re tools the team uses every day.”source
“At critical junctures he navigated the team to success using a mixture of technical knowledge, pragmatism, and a solution based mindset.”source