| name | writing-agent-managed-skills |
| description | TraceDecay Dev: Use when creating, revising, validating, approving, installing, or auditing TraceDecay agent-managed automation skills and skill-writer drafts. |
TraceDecay Dev: Writing Agent-Managed Skills
Agent-managed skills are profile-owned drafts produced by automation. Treat
them as generated artifacts with lifecycle state, validation evidence, and
usage telemetry; bundled plugin/skills/ changes are a separate release path.
Workflow
- Inspect before editing:
tracedecay_skill_list for inventory, then
tracedecay_skill_view with support files for the target draft.
- Read the writer run evidence:
tracedecay_automation_run_artifact_view for generated_evals,
validation_gate, optimizer_diagnosis, and codex_handoff artifacts.
- Decide the path:
- managed draft fix: update through
tracedecay automation skills ...
commands or dashboard flow;
- bundled reusable guidance: edit
plugin/skills/<slug>/SKILL.md and run
agent-suite validation.
- Validate discovery text: description starts with
Use when, names concrete
triggers, avoids workflow summaries, and stays host-portable.
- Check adoption after install with analytics diagnostics and skill usage
events before declaring the skill effective.
Quality Bar
| Check | Pass condition |
|---|
| Trigger | Another agent can identify when to load it from description alone. |
| Body | Short imperative workflow, no session narrative, no stale environment facts. |
| Evidence | Run artifact or transcript shows the failure this skill prevents. |
| Validation | Draft has a validation gate or bundled skill passes plugin tests. |
| Adoption | Usage telemetry or follow-up session proves it was invoked. |
Guardrails
- Never approve or install a managed skill without explicit user intent.
- Never copy managed-skill state into bundled skills unless the pattern is
reusable across projects.
- Do not mutate profile stores from subagents. Subagents may inspect and
recommend only.
Deliverable
Return the skill id or bundled path, lifecycle state, validation artifacts read,
edits made or recommended, adoption evidence, and the exact approve/install
command when mutation is appropriate.