Skip to content

Skills and agents

Beyond the pipeline commands, the workspace defines skills (domain guides Claude applies when working in an area), role personas (bundles of skills per project type), and uses sub-agents for autonomous, parallelizable work.

Task skills

Domain-specific guides in .claude/skills/. Reference them when working in the matching area.

Skill Use for
code-reviewer Structured code review with checklists and severity ratings
code-architect System design, ADRs, project structure, scalability
code-creator Scaffolding, boilerplate generation, feature implementation
code-validator Linting, type checking, testing, static-analysis pipeline
web-developer Frontend/backend, responsive design, full-stack integration
database-engineer Schema design, query optimization, safe migrations
devops-engineer Docker, CI/CD, deployment, monitoring
api-designer REST conventions, OpenAPI specs, API security
security-auditor OWASP Top 10 audit, vulnerability assessment
performance-optimizer Profiling, caching, bottleneck resolution

Role personas

Personas bundle the task-skills appropriate to a project type. They are routed per repo via project-skills plus each .claude/<repo>/CLAUDE.md.

Persona Bundles for
backend-developer Service/API backend repos
data-engineer Data pipelines / ETL
integration-api External integration / provider APIs
infra-devops Infrastructure and deployment repos
library-maintainer Shared libraries / packages
frontend-developer UI repos (reserved for a future UI repo)

Meta / routing skills

Skill Role
project-skills Per-repo routing: which persona + skills apply to which project
task-creation-pipeline The end-to-end task-delivery procedure (source of truth for templates, statuses, and the mock -> real Jira mapping)
model-routing Per-phase Claude tier selection (Opus / Sonnet / Haiku) for token/cost optimization
context-cache Mechanics of the pre-summarized, token-saving context cache

Agents (sub-agents)

Heavy, autonomous, artifact-producing work is delegated to sub-agents via the Agent tool, often on a cheaper model tier (see model routing). Sub-agents cannot prompt the user, so interactive approval gates always stay on the main session.

Typical uses:

  • Implementation during /approve-spec -- delegated to a coding agent (Sonnet, escalated to Opus for large/high-risk specs).
  • Broad search / research -- an exploration agent sweeps many files and returns only the conclusion, keeping the main context lean.
  • Parallel independent work -- multiple agents run concurrently when tasks don't depend on each other.

Skills vs commands vs rules

Commands are things you invoke (/name). Skills are guides Claude applies while working in a domain. Rules apply automatically on every session (next page). All three are plain markdown under .claude/.