Skills and Automated Workflows#

The difference between using Claude Code conversationally and using it effectively is skills. A skill is a structured instruction set that tells Claude exactly how to approach a specific type of work on a project. Instead of explaining your project’s conventions every session, you encode them once and invoke them by name.

For how skills work technically, see the official skills documentation.

Our approach: skills as institutional knowledge#

We treat skills as the place where project-specific knowledge lives. A good skill encodes:

  • Project conventions — the architecture patterns, naming rules, and file organization that are specific to this project
  • Past mistakes — things that went wrong and what the skill should do differently
  • Quality gates — verification steps that must pass before the skill considers its work done

Skills load on demand, not on every message. This makes them a better place for detailed workflows than CLAUDE.md, which loads every time.

Iteration logs#

The most impactful practice we’ve found is maintaining iteration logs inside skills. After every use of a skill, record what went well, what broke, and what to do differently next time. The skill evolves with every use — Claude doesn’t repeat mistakes because the fix is encoded in the skill itself.

Teams on our PRM project saw Figma-to-code workflows improve significantly over successive iterations: each round of learnings reduced redundant API calls, improved component reuse, and eliminated repeated errors.

Examples from our projects#

Drupal SDLC Plugin#

The Drupal SDLC Plugin is a Claude Code plugin built for Drupal projects at Axelerant. It ships with 11 skills covering the full development lifecycle:

  • /work-on-jira-ticket — reads a Jira ticket, generates a spec, builds, validates, tests, and raises a PR
  • /spec-writer — converts a ticket or description into a structured spec with acceptance criteria and test mapping
  • /config-builder — generates Drupal config YAML from natural language
  • /validate — runs the complete quality gate: PHP lint, GrumPHP, config round-trip, visual screenshots, tests
  • /pr-reviewer — reviews a PR diff for security, AC coverage, and risk

The plugin also includes hooks that block writes to protected files and run PHP syntax checks on every edit. See the full documentation on Confluence for setup and usage.

Project-specific skill catalogs#

Several teams have built skill catalogs tailored to their projects:

  • PRM (Next.js + Magento) — 10 skills covering Figma-to-UI, API integration, unit and e2e test generation, ticket-to-PR automation, and Jira task creation
  • CDM teams — skills for Pattern Lab scaffolding, Drupal integration, implementation details generation, and structured code review

These catalogs are shared in #wg-ai-native-engineering on Slack. If you build skills that could benefit other teams, share them there.

Building your own skills#

Start with the tasks your team repeats most often. A good first skill is usually one that:

  • Follows a consistent sequence of steps every time
  • Requires project-specific knowledge that Claude wouldn’t know on its own
  • Has clear verification criteria (tests pass, lint is clean, config is valid)

The official documentation covers the technical format. What matters more is the content: be specific about your project’s patterns, include the mistakes you’ve learned from, and define what “done” looks like.