Agent skills are blowing up. I wanted one that actually helps: score the repo, show evidence, then only change stuff after you say so.
Vibe Cleaner does that. It measures first (deps, artifacts, code health, verification), spits out a 0–100 score and two artifacts:
cleanliness_score.json and cleanliness_report.md. Changes go through two stages. Stage 1 is safe removals (unused deps, dead files, generated junk).
Stage 2 is structural reorg (moves, import path changes) and only runs after Stage 1 is done and you’ve verified.
So “clean up” doesn’t turn into “oops we moved everything.”
flowchart TD
A[Trigger: local / CI / chat] --> B[Inventory & profile]
B --> C[Collect signals: deps, artifacts, code health, verify]
C --> D[Compute score + write cleanliness_score.json & cleanliness_report.md]
D --> E{Decision gate}
E -->|Score ≥ 85| F[Optional targeted cleanup]
E -->|Stage 1 needed| G[Stage 1 proposal: safe removals with evidence]
G --> H{Approve Stage 1?}
H -->|No| Z[Done]
H -->|Yes| I[Apply Stage 1: remove deps / dead files / artifacts]
I --> J[Verify: lint, typecheck, tests]
J --> K[Recompute score + update reports]
K --> L{Stage 2 needed?}
L -->|No| Z
L -->|Yes| M[Stage 2 plan: moves, splits, import path changes]
M --> N{Approve Stage 2?}
N -->|No| Z
N -->|Yes| O[Apply Stage 2]
O --> P[Verify again]
P --> Z
F --> ZYou may visualize it on Mermaid Web
Config is optional (.vibe-cleaner.json in the repo root). The skill ships with scripts so you can run the same pipeline locally or in CI.
You trigger it by what you say. “Clean up this repo,” “Is this codebase clean enough to refactor?” it scores, reports, then asks for approval before touching anything.
Get the skill and full docs here: [SKILL_LINK].