AI Skills & Workflow
Most AI workflows are just better autocomplete. Mine isn't. I treat prompts the way I treat any repeatable system: if I'm writing the same thing twice, I abstract it. Skills replace repetition. Agents handle the boring parts. The result is a workflow where AI manages the process, so I can focus on the thinking that requires context, not computation.
The tools change every month. What lasts is the system around them: the context AI gets, the workflows that connect it to real users, and the review that decides what ships. I share what works, and what doesn't, with the rest of the team.
Workflows
AI only builds things that look and work like the product when it has the right context. At XYCLE I wrote the skills, codebase rules and design system context that give AI our real components, tokens and constraints, using Figma MCP and Claude Design as tools. Figma, Storybook and production stay aligned across 40+ components, so everyone's output gets better, not just mine.
Customer feedback used to live in call recordings nobody had time to rewatch. I set up Grok bots connected to Fireflies, Granola and PostHog that turn customer calls into tickets, alert the team when PostHog shows people getting stuck, generate reports, and surface UX gaps and feature ideas for product, engineering and customer success.
AI is fast and confidently wrong. It invents components, skips edge cases and misses accessibility. Every AI-generated change goes through the same code review, tests and design QA as anything else, and recurring mistakes become new rules in the codebase so they don't happen twice.
Planning
Manus-style file-based planning for complex multi-step tasks. Creates task_plan.md, findings.md, and progress.md, with hooks that survive /clear for automatic session recovery.
Synthesises planning files into a structured prd.md. Run after planning-with-files to turn raw research into a shippable spec, no questions asked.
Breaks prd.md into independently-grabbable vertical-slice tasks with acceptance criteria, dependencies, and suggested order. Output is a single tasks.md ready for tasks-to-jira.
Creates Jira issues from tasks.md via the Atlassian MCP. Handles story points, issue types, epics, and native blocking links by creating issues in dependency order.
Prototyping
Before committing engineering resources, I prototype. With Claude Design I go from concept to interactive prototype in a single conversation, connected to the real codebase, using actual components, aware of existing patterns. Fast enough to validate the same day, specific enough that the handoff to production is short.
claude.ai/designDevelopment
Once I'm building, tdd keeps me honest. Red, green, refactor, one vertical slice at a time.
When the work is ready to ship, pr-description writes it for me. Generated from the git diff and branch name, with the Jira link, visual changes, and test steps already filled in.
Set up Husky pre-commit hooks with lint-staged, Prettier, type checking, and tests.
Design intelligence
When I join a new product or start a design system from scratch, I don't want to start from zero. skillui points at any live URL, repo, or folder and extracts the full system: colours, typography, spacing, components, animations, packaged as a .skill file. The insight is still mine. The inventory isn't.