I run 9 Always-On agents with Claude Code and claude-code-hermit plugin. These are the 4 official Anthropic plugins I run on all of them.
claude-code-setup
I run it as a weekly routine
Analyzes the repo (stack, structure, package.json, layout) and recommends the highest-value automations to add: skills, hooks, MCP servers, subagents, slash commands.
Skill:
/claude-code-setup:claude-automation-recommender
claude-md-management
Run the audit weekly or the revise-md on session-close or on a reflection schedule
Audit and improve CLAUDE.md files in repositories. Scans for all CLAUDE.md files, evaluates quality against templates, outputs quality report, then makes targeted updates. There's also a revise-claude-md skill where the session context is used for revision which is really useful in reflection routines in always-on agents.
Skills:
/claude-md-management:claude-md-improver # general audit on your CLAUDE.md files
/claude-md-management:revise-claude-md # revise with session context
claude-md-management on GitHub →
skill-creator
Whenever I want to create, edit, evaluate, benchmark, and improve skills
Use when you want to create a skill from scratch, edit, or optimize an existing skill, run evals to test a skill, benchmark skill performance with variance analysis, or optimize a skill's description for better triggering accuracy.
Skill:
/skill-creator:skill-creator [describe the skill you want to create, modify or improve]
feature-dev
For the changes big enough to deserve a plan instead of a one-shot edit.
The other three recommend, maintain, and package. feature-dev builds. It runs a seven-phase workflow (discovery, exploration, clarifying questions, architecture, implementation, review, summary) backed by three read-only Sonnet agents: code-architect blueprints the files, code-explorer traces execution paths, code-reviewer flags only high-confidence issues. Crucially, it gates on your approval before writing any code and offers architecture options first, a checkpoint that matters more, not less, when the agent runs with some independence. Not for one-line fixes or hotfixes; for the work that deserves a plan without entering Plan mode.
Skill:
/feature-dev:feature-dev [optional feature description]
Why these four
Recommend, maintain, skill, build. That's the loop.
Top comments (0)