V3.5 adds /create-agent — an interactive Q&A command that builds a custom
agent tailored to your project’s domain. Pick a template, name it, specialise it, set
a tone, pick tools, preview the result before writing anything.
Custom agents use a custom- prefix so /update-reviewsquad never
overwrites them, and /quick dispatches them directly alongside squad agents.
-
5-question interactive Q&A
The command walks through five structured questions: (1) template selection, (2) agent name, (3) specialization focus, (4) tone, (5) tool selection. Each answer shapes the agent prompt — the output is a ready-to-use agent file, not a generic shell.
-
6 starter templates
Security reviewer — OWASP, threat modelling, auth flows. Code quality — patterns, conventions, readability. Domain expert — project-specific knowledge (e.g. PCI DSS, ML pipelines). Documentation — docstrings, API docs, comment quality. Performance — profiling, query analysis, runtime complexity. Blank — start from scratch.
-
Preview before write
After answering all five questions, the command displays the full agent definition as it would be written. You confirm before anything is saved to disk. If the output isn't right, the session can be restarted — nothing is committed until you say yes.
-
custom- prefix — never overwritten by /update-reviewsquad
All custom agents are saved to ~/.claude/agents/custom-{name}.md. The /update-reviewsquad command explicitly skips any file matching custom-*.md. Your domain-specific agents survive every squad update, regardless of which files changed upstream.
-
Dispatched via
/quick like any squad agent
Custom agents integrate into the existing /quick dispatch system. Use /quick <task> custom-{name} to run a custom agent on any task, or mix custom and squad agents in the same invocation: /quick <task> custom-payments-expert,jared:review.
-
Tone and tool configuration
Four tone options: direct/blunt (mimics Jared), collaborative, formal, or neutral. Tool selection defaults from the chosen template but can be overridden — a documentation agent typically only needs Read and Write; a performance agent might want Bash for profiling commands. The configuration is captured in the agent prompt, not in YAML frontmatter.