Reusable Meta_Kim file inventory classification flow. It helps separate durable sources, generated evidence, runtime mirrors, temporary state, and risky unknowns before cleanup or commit.
Installs into .claude/skills of the current project.
Are you the author of Same Set Reusable Flow For Project File Inventor?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/kimyx0207-same-set-reusable-flow-for-project-file-inventor)
---
name: same-set-reusable-flow-for-project-file-inventor
version: 0.1.0
author: Meta_Kim
user-invocable: false
candidateType: skill
sourceGapId: gap-create-skill-same-set-reusable-flow--c89068df05
approvalEvidence: warden-approved-file-inventory-skill-2026-06-05
trigger: "project file inventory|file classification|repo cleanup|项目文件整理|文件用途整理|清理项目"
tools:
- shell
- filesystem
description: |
Reusable Meta_Kim file inventory classification flow. It helps separate durable sources, generated evidence, runtime mirrors, temporary state, and risky unknowns before cleanup or commit.
---
# Project File Inventory Classification
Use this skill when a Meta_Kim workspace has too many files and the user needs to know which are useful, generated, risky, or safe to ignore.
## Goal
Produce a source-backed inventory, not a cleanup guess. The output should help the user decide what to keep, commit, ignore, or inspect again.
## Required Evidence
Start with read-only evidence:
1. `git status --short --branch`
2. `git ls-files`
3. `git ls-files --others --exclude-standard`
4. `git check-ignore -v <path>` for suspicious ignored paths
5. `rg -n "<filename-or-command>" package.json docs scripts tests canonical config` for tracked candidates
6. `npm run` script inventory from `package.json`
Do not classify a tracked file as removable only because it has few references. Check package scripts, tests, docs, sync scripts, setup scripts, and contracts first.
## Classification Buckets
- **Core source**: `canonical/`, `config/contracts/`, `config/capability-index/`, durable scripts, package scripts, and tests that define behavior.
- **Runtime projection**: `.claude/`, `.agents/`, `.cursor/`, `openclaw/`, `.codex/` mirrors generated from canonical sources or sync scripts.
- **Generated evidence**: `.meta-kim/state/default/`, graph outputs, reports, SQLite run state, smoke artifacts, and temporary proof files.
- **Planning continuity**: `task_plan.md`, `findings.md`, `progress.md`; keep during active work even if ignored.
- **Reference docs**: tracked docs and plans that explain product, architecture, contracts, or release decisions.
- **Suspicious or cleanup candidate**: files with no package, test, docs, sync, install, contract, or runtime evidence.
## Output Format
Return a compact table with:
- path or pattern
- bucket
- why it exists
- evidence command
- recommendation: keep, commit, ignore, regenerate, inspect again, or cleanup candidate
- risk if deleted
End with a short "do not delete yet" list for any file whose role is uncertain.
## Boundaries
- Do not auto-write canonical state without Warden approval.
- Keep one-run task details in workerTaskPackets, not durable identity.
- Do not delete files as part of inventory classification unless the user separately approves a cleanup action.
- Do not treat runtime mirrors or generated reports as source of truth when canonical or config sources exist.