SuperJunior is a collection of skills for coding agents. Use it to carry a code change from scope to verification, or to understand a system before deciding what to build.
Choose the result you need:
| Your task | Start with |
|---|---|
| Add a feature, fix a defect, or refactor | junior-mode |
| Explain code, investigate a design decision, or compare architectures | senior-mode |
| Scope, implement, debug, review, or commit one bounded piece of work | The matching agent role |
Junior mode coordinates implementation and checks the evidence from each phase. Senior mode develops explanations and decisions without changing product code. Individual roles also work directly in the current session; using them does not automatically start subagents.
Copy the directories inside skills/ into the target project's skill directory, keeping them side by side. The project's documented locations are .claude/skills/ for Claude Code and .agents/skills/ for Codex CLI and Antigravity CLI. Use ordinary directory copies so the links between skills remain intact.
See Host integration for the recorded discovery baseline, personal directories, and troubleshooting.
In your coding agent, try:
Use junior-mode to fix the saved filter.
With no selected tags, it must show all items.
Keep the behavior for selected tags unchanged.
Reproduce the failure and verify the fix.
Leave the changes in my working tree without committing.
State the delivery you want: working-tree changes, local commits, or another explicitly requested action. Installing a skill does not grant tools or permissions.
Each role reports evidence in a form suited to its task. A debugger's IDENTIFIED describes a diagnosis; a reviewer's APPROVE describes a review. Junior mode checks those results before declaring the requested work complete.
Start with the handbook, follow the first-task walkthrough, or adapt an example request.
The execution instructions live in skills/<name>/SKILL.md; their user-facing explanations live in docs/. Keep both accurate when behavior changes.
Run npm ci once, then npm run check to build the handbook and test its links, skill definitions, eval graders, and workspace checks. These local tests do not call an agent. Maintenance covers local preview, CLI evals, and publishing.
MIT. See LICENSE.