superskill

SuperSkill

SuperSkill is a consumer skill for a host-configured superskill MCP server, not a standalone API or offline service. Use only tool names actually exposed by the host: OpenCode Code Mode uses tools.superskill.<tool>, while Claude Code exposes mcp__superskill__<tool>. Other hosts may differ; never guess or mechanically rewrite a tool name. If a tool is unavailable, report that or use an already-installed CLI after verifying it with superskill --help. Do not install tools or use npx unless network use is explicitly approved.

Scope and authorization

• Treat a coding request as authorization for its specified project and changes only. Establish the actual checkout and project; never guess a project slug. If mapping is needed for authorized work, call register with the absolute checkout path and use its returned slug.
• Session registration and completion are coordination bookkeeping for authorized nontrivial code work; completion persists a session note. This does not authorize additional ADRs, learnings, tasks, or reports unless requested or required by project policy.
• Keep data scoped to the named project; access other project history only with explicit scope and authorization. Redact prompt text by default; use include_prompts only for bounded excerpts from the current project when explicitly requested and needed. Never expose secrets, credentials, or unrelated history.
• Do not infer permission for network access, installs, global configuration, fixes, dismissals, manual cleanup/cache apply/GC, publishing, or destructive changes. The built-in session-maintenance preservation path below is the exception; preview manual actions and proceed only with explicit authorization.

Workflow

• Use the smallest workflow that fits. Simple questions and read-only lookups need no session, plan, HLD/LLD, or human-question ritual; ask only when scope, authorization, or risk is genuinely unclear.
• For nontrivial authorized code work, load project_context and search relevant project notes before planning. Register once with the session tool using action: "register", tool, project, and absolute workspace_path; retain session_id.
• Route with superskill using task, repo-relative files, session_id, workspace_path, and phase (explore|implement|review|ship). Reuse the same session; reroute when task, files, or phase meaningfully change. Complete that session with the outcome and verification evidence.
• Inspect the returned worktree assessment. Use worktree_env before installs/builds. Built-in session maintenance may automatically quarantine or relocate eligible clean, idle, session-owned checkouts after session completion once they are proven integrated into the repository's actual default branch; it preserves the whole checkout in a locked recovery archive, returns recovery metadata, and never deletes it. An assessment alone does not authorize manual actions. Manual cleanup, cache apply/GC, permanent removal, policy/config changes, and installs require explicit authorization; never force-remove a checkout.