30 skills and 4 agents that take a Jira ticket from idea to verified PR, with humans in charge.

The problem

The team was moving from AI-assisted to AI-delivered work, and needed one consistent, verifiable way to take tickets to PRs with humans keeping the decisions.

What we did

• Forked the Travelopia DX workflow and rebuilt it on Jira and the Yarn, Prisma and Turborepo stack.
• Chained skills from shaping to spec to tickets to test-first builds in isolated worktrees, with Jira status and a Report comment updated automatically.
• Built /prove-it: a fresh verifier agent that did not write the code tests each PR in a real browser against a checklist for roles, payments, PII, accessibility and OWASP.
• Added retros on every run and a fortnightly gardening pass that turns any lesson seen twice into a type, lint rule, test or CI check.
• Set guardrails: agents never merge, close tickets or deploy, and stop for approval before CDK stacks, data-dropping migrations and .env files.

Key decisions

1. 01

   A separate verifier agent that never grades its own work

   Trade-off: More time and tokens per PR, in exchange for verdicts reviewers can trust.

2. 02

   Humans always merge and deploy

   Trade-off: Slower than full automation, but accountability stays with people.

3. 03

   Route repeated lessons to the strongest guard first

   Trade-off: More upfront work per lesson, so mistakes stop repeating instead of being re-explained.

Architecture

Result

Tickets move from idea to verified PR the same way for every developer, and repeated mistakes become permanent checks instead of review comments.

What I learned

1. 01

   Agents need the same context a new hire needs: a glossary, ADRs and a map of the code.

2. 02

   The guardrail that matters most is who is allowed to say "done".

Stack

• Claude Code
• Skills
• Subagents
• Hooks
• Git Worktrees
• Jira
• Playwright
• GitHub Actions