| name | as |
| description | Run any registered role from the repo's evolving role repository (.claude/roles/) against a target — the generic solo dispatcher for roles that have no dedicated wrapper skill. Invoke explicitly. |
| argument-hint | [role] [target] |
| disable-model-invocation | true |
Run a role solo
Input: $ARGUMENTS — the first word is the role name; the rest is the target.
- Resolve the role. Prefer the repo's instantiated copy
.claude/roles/<role>.md (the
canonical, evolving one; registry.md indexes these). If it's absent, check this plugin's
bundled seeds alongside this skill at ../<role>/role.md — if found, copy it verbatim to
.claude/roles/<role>.md and use that. If the role exists in neither, list the available roles
(repo .claude/roles/*.md plus the bundled ../*/role.md, excluding crew.md/panel.md/
registry.md/this as skill) and stop — don't improvise a persona.
- Apply it. Adopt the role's Charter and Body; honor every entry under
## Learnings (core)
and ## Learnings (solo). Note crew.md and panel.md are consumer registries, not roles —
never load them as personas.
2.5 Anchor in the Prompt Library (optional). Each engineering persona maps to a category of
Anthropic's Claude Code Prompt Library —
architect→Plan, builder→Implement, debugger→Debug, optimizer→Debug, refactorer→Refactor,
reviewer→Review, skeptic→Review (the role's Body also names its category). If the
prompt-coach-beta plugin is installed, run its config.py library --category <Category> for
gold-standard prompt shapes in this role's domain and let them inform your opening. Degrade
silently — this is enrichment, not a dependency; skip it if the plugin/command isn't present.
- Work the target following the Body's method and deliverables.
- Capture a learning. If the run surfaced a durable, reusable lesson (a fix that took more
than one cycle, a stack- or repo-specific gotcha, a stated user preference), append one line
under
## Learnings (solo) in the role file: - YYYY-MM-DD — <one sentence>. Don't log
routine one-pass outcomes.
Charter and Body edits are never made during a run — propose them to the user instead
(identity changes are deliberate, user-gated).
