| name | workflow-evaluator |
| description | Use this skill when evaluating whether Ravie skills, workflows, routing, or approval gates are working after real usage, recurring friction, skipped gates, or a monthly health check. Do not use for product feature work or to blame a skill without evidence from actual sessions. |
workflow-evaluator
Evaluate whether skills and workflows are actually working. Find the skills that route incorrectly, produce bad output, or get bypassed.
Purpose
The quality check for the system itself. After real usage, evaluate which skills helped, which produced wrong output, which got ignored, and which gates got bypassed.
When to use this
- After
pattern-learner identifies a skill that produced bad output
- When a permission gate was skipped during a session
- When a workflow felt slow or wrong
- Monthly system health check
When NOT to use this
- Do not use for product feature work or to blame a skill without evidence from actual sessions.
Process
1. Collect evidence
For each skill being evaluated:
- How many times was it invoked this week/month?
- Did it produce output that was used as-is, or required heavy editing?
- Did it route correctly (triggered for the right tasks)?
- Did it miss triggers (should have activated but didn't)?
- Were approval gates respected?
2. Classify skill health
| Status | Meaning | Action |
|---|
| Healthy | Triggers correctly, output is usable | No change needed |
| Noisy | Triggers too often for irrelevant tasks | Tighten trigger conditions |
| Silent | Should trigger but doesn't | Broaden trigger conditions or update router |
| Stale | Output references outdated patterns | Update the skill content |
| Wrong | Produces incorrect or harmful output | Rewrite or remove |
| Unused | Never triggered in 4+ weeks | Evaluate if still needed |
3. Check gate compliance
Review Tier 3-4 actions from the past week:
- Were approval gates respected?
- Were any Tier 4 actions blanket-approved (violation)?
- Were any actions misclassified to a lower tier?
4. Propose fixes
For each unhealthy skill:
- Specific change to make
- Evidence for why
- Expected improvement
Output format
# Workflow evaluation — [date range]
## Skill health
| Skill | Status | Invocations | Notes |
|---|---|---|---|
| issue-to-pr | Healthy | 12 | Output used as-is |
| deploy-ready | Stale | 3 | References outdated Vercel config |
| figma-lovable-handoff | Silent | 0 | Should have triggered for UI work |
## Gate compliance
- Tier 3 actions: [N] total, [N] properly gated
- Tier 4 actions: [N] total, [N] properly gated
- Violations: [list]
## Proposed changes
- [ ] [Skill]: [change] — because [evidence]
## Skills to consider removing
- [Skill]: unused for [N] weeks, no clear use case remaining
Hard rules
- Never evaluate based on theory — only based on observed usage
- Never remove a skill without checking whether it was silent (should have triggered but didn't)
- Never ignore gate violations — they indicate a structural problem
- Always cite specific evidence for proposed changes
Connects to
pattern-learner — feeds findings into the weekly review
skill-creator — for creating skills to fill gaps identified here
permission-guardian — for reviewing gate compliance
Common failure modes
Blaming the skill when the router is wrong — A skill produces bad output because it was triggered for the wrong task. Check the routing logic, not just the skill.
Removing unused skills prematurely — A skill might be unused because it's rare, not because it's useless. Check whether situations where it should trigger actually occurred.