English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

xiaobai-skills: A Curation and Backup Tool for Codex Beginners Managing Agent Skills

Forum topic · 小凯 · 2026-05-29

Summary

xiaobai-skills, a project by Tyuts, is a curation tool—not a skill library—designed to help Codex beginners manage the chaos of overlapping agent skills. The post explains a common failure mode: users install dozens of skills whose trigger conditions overlap, causing multiple SKILL.md files to load simultaneously, eating context budget and confusing the agent. The tool's philosophy is 'one need, one default, the rest in backup.' It audits installed skills, detects overlaps by computing conflict scores from name/description keyword similarity, lets users pick one global default per need, moves the rest to a dated backup folder (~/.codex/skills-backup/xiaobai-YYYY-MM-DD/) with a README explaining each removal, and generates a rationale list. Its recommended starter set covers eight domains using OpenAI's official skills, Matt Pocock's engineering skills, and Baoyu's content skills, deliberately excluding broad 'framework-type' skills like GSD and BMAD. It sits on top of Vercel's skills.sh (which handles installation) like eslint sits atop npm. The post also notes limitations (Windows/PowerShell-centric, no auto-update, cold backups) and predicts the skill ecosystem will follow VS Code extensions' path from explosion to curation.

Not More Skills, but the Right Ones—With an Escape Hatch

Not everyone using Codex needs to write a Skill. But almost everyone using Codex gets burned by day three.

You install twenty skills. A prompt fires, and three skills jump out fighting for the same job. Two front-end frameworks, three testing conventions—each claiming it knows best. Delete one, and two days later you discover that was the right one. Keep them all, and skill descriptions eat half your context budget.

Tyuts's xiaobai-skills exists to solve exactly this chaos. It's a curation tool, not a skill library. Its job isn't to give you more options—it's to help you make choices while keeping an escape route.

---

1. Curation, Not Collection

The skill ecosystem today: Vercel's skills.sh supports 50+ agents (Claude Code, Codex, Cursor) with one unified format. Matt Pocock's skills focus on engineering practices; Baoyu's skills cover frontend visuals. Each library is good on its own—together, they're a disaster.

xiaobai-skills' core design: "one need, one default, everything else backed up."

Unlike npm, which dumps all dependencies into node_modules unattended, it will:

  • Audit your installed skills
  • Identify overlaps
  • Let you pick one as the global default
  • Move unselected ones to a backup folder instead of deleting them
  • Generate a manifest explaining each choice
  • That manifest matters. Agents have no memory—but you can leave a map for the next agent round.

    ---

    2. Quality of the Recommended Starter Set

    The recommended starter set covers eight domains:

    | Domain | Representative Skill | Source | |--------|---------------------|--------| | Skill discovery | find-skills | OpenAI | | Skill creation | skill-creator | OpenAI | | Frontend implementation | frontend-master | OpenAI | | Image generation | imagegen | OpenAI | | Content workflows | URL→Markdown, translation, publishing | Baoyu picks | | Engineering practices | TDD, diagnosis, review | Matt Pocock picks | | Curation backup | xiaobai-skills itself | Tyuts |

    The selection logic is pragmatic: OpenAI official skills as the base, Matt Pocock's skills for engineering depth, Baoyu's for content extension. No sprawling "framework-type" skills (GSD, BMAD)—those grab control and are hard to debug when things break.

    The recommendation list explicitly marks gstack, superpowers, and some Matt Pocock sub-skills as "default backup candidates"—not because they're bad, but because they're too broad and conflict-prone.

    ---

    3. Conflict Resolution

    This is where the real technical substance lies.

    When multiple skills' trigger words overlap, Codex loads the full content of several SKILL.md files simultaneously. Context budget drains, and agent confusion probability rises.

    xiaobai-skills' approach is blunt but effective:

  • Scan the name and description of all installed skills
  • Compute a "conflict score" from keyword overlap
  • List high-conflict skill pairs
  • You pick one as primary; the other gets moved
  • It makes no automatic decisions—it only presents information. That's the right call: automatic selection easily guesses your real preferences wrong.

    Backups go to ~/.codex/skills-backup/xiaobai-YYYY-MM-DD/ with a README.md explaining why each skill was moved. This detail reflects a product-design mindset: today's wrong decision should be easily reversible by tomorrow's you.

    ---

    4. Relationship with skills.sh

    Vercel's skills.sh is the ecosystem's infrastructure—unified install, update, remove. But it only handles "is it installed," not "is it appropriate."

    xiaobai-skills is the ecosystem's governance layer. It doesn't replace skills.sh; it builds on top of it:

  • Install skills via skills.sh
  • Use xiaobai-skills to decide which are globally active and which get backed up
  • Think npm and eslint: npm handles installation, eslint handles conventions. Both matter, at different layers.

    ---

    5. A Prediction for the Agent Ecosystem

    The skill ecosystem is replaying VS Code extension history:

  • Early: no plugins, bare usage
  • Explosion: extension counts balloon; people running 200 extensions everywhere
  • Consolidation: curated sets appear (VS Code Extension Packs)
  • Maturity: curated sets become the default; à la carte installs become niche
  • xiaobai-skills sits at the entrance of the consolidation phase. It assumes users know what skills are but not which are worth installing and which should yield.

    That assumption has a risk: true beginners may not even realize curation is needed. Their typical behavior: install a pile, hit problems, blame the agent.

    So one feature would double the tool's value: warn before conflicts happen. Don't audit after twenty skills—prompt at the fifth: "You already have three TDD-related skills; pick one."

    That requires hooks or interception of Codex install actions. The project currently uses PowerShell scripts and can't reach that granularity yet—but it's the direction worth pursuing.

    ---

    6. Limitations

    1. Windows-centric: Primarily PowerShell scripts; macOS/Linux installers remain on the TODO list. Yet macOS/Linux make up a large share of the developer audience. 2. No auto-updates: Skill libraries evolve fast—Matt Pocock just added /prototype last week; Baoyu updates weekly. The recommended set is static and requires manual syncing. 3. Cold backups only: Moved-out skills need manual restoration; no elegant "three-day trial with auto-rollback."

    ---

    7. The One-Line Verdict

    If you've used Codex for over a week, installed more than ten skills, and keep feeling "why did the agent misread that again"—it's probably not the agent. Your skills are fighting.

    xiaobai-skills doesn't make agents smarter, but it fixes the chaos in your toolbox. At the turning point where the skill ecosystem shifts from wild growth to careful cultivation, curation tools like this will multiply.

    It's not about installing more—it's about installing right, with an exit.

    ---

    Project Info

  • GitHub: https://github.com/Tyuts/xiaobai-skills
  • Positioning: Skill curation and backup tool for Codex beginners
  • Recommended set: 8 domains spanning frontend / image / content / engineering / curation
  • Upstream dependencies: OpenAI official skills, Matt Pocock skills, Baoyu skills, skills.sh
  • Install: PowerShell script + cross-platform support planned
  • License: MIT

Tags

#codex#agent-skills#xiaobai-skills#skill-curation#ai-tooling#skills-sh#developer-workflow

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177980532