Skills

NVIDIA SkillSpector audit AI review by Claude

A small collection of practical Agent Skills.

Public monorepo. Anyone can clone this repo, run just add, or install individual skills with npx skills add. It does not depend on any private tooling. The install loop links into Claude Code, Codex, and Grok skill dirs.

Catalog: self-authored skills live under skills/ (see also the GitHub Pages registry). Highlighted ones below; the full set is every folder with a SKILL.md. Third-party modules are declared in external.yml and materialised by just vendor.

Skills

claude-5-best-practice

Renamed from fable-5-best-practice in July 2026, when Claude Opus 5 shipped near-Fable capability at half the price and the routing advice stopped being about one model. Reinstall under the new name.

Optimizes the cost of an accepted result across the Claude 5 generation — Haiku 4.5 / Sonnet 5 / Opus 5 / Fable 5 — rather than the price of a token. It treats Opus 5 as the default lead, makes Fable 5 something you justify instead of assume, and prefers climbing the effort ladder (low → max) on Opus 5 before escalating a tier. It also walks a diagnose-before-escalating loop for failures, routes subagents with pinned worker models, covers Claude 5 API behavior (adaptive thinking, effort and prompt caching, refusal stop reasons, fallback, ZDR and Priority Tier constraints), and keeps progress claims tied to real tool evidence. Volatile facts such as pricing, retention, and beta headers live in a dated references/evidence-notes.md provenance snapshot.

The skill is especially useful when migrating older prompts or skills forward: it calls out over-constrained legacy prompting, the Opus 4.8 → Opus 5 thinking and effort breaks, refusal risks, verifier agents, long-running run design, and long-run communication style.

Install:

npx skills add https://github.com/wquguru/skills --skill claude-5-best-practice

Ask for help with prompts like:

  • "Does this task actually need Fable 5, or is Opus 5 enough?"
  • "Help me write an Opus 5 prompt for a multi-day coding task."
  • "Review this old skill before migrating to Claude 5."
  • "My run missed the acceptance criteria — raise effort or change model?"

gpt-5-6-best-practice

A routing and evaluation overlay for GPT-5.6 in Codex. It optimizes the cost of an accepted result — not the price of a token — across the Sol, Terra, and Luna capability tiers: choosing the initial tier and reasoning-effort lane, diagnosing failures before escalating, keeping prompts and worker packets lean, using subagents economically, and evaluating routes on real work. Benchmark provenance and confidence levels live in references/evidence-notes.md, and references/prompt-patterns.md provides reusable task-prompt and worker-packet templates.

Install:

npx skills add https://github.com/wquguru/skills --skill gpt-5-6-best-practice

Ask for help with prompts like:

  • "Which GPT-5.6 tier and effort should this task start on?"
  • "This Codex run failed — raise effort, or move from Terra to Sol?"
  • "Is Ultra worth it for this task, or should one agent handle it?"
  • "Design a worker packet for this subagent."

english-swe-daily

A practical English coach for everyday engineering communication. Built for developers who can already work in English but want to sound less textbook-like and more like a real teammate in standups, Slack threads, 1:1s, code reviews, and meetings.

Many engineers do not struggle with grammar — they struggle with sounding natural. This skill teaches the kind of English people actually use in software teams, with a strong focus on natural phrasing, collaboration tone, and day-to-day workplace situations.

Install:

npx skills add https://github.com/wquguru/skills --skill english-swe-daily

Ask for help with prompts like:

  • "Teach me English for standups."
  • "Help me sound more natural in Slack."
  • "Practice English for code reviews."
  • "How do I say this more naturally at work?"
  • "Give me today's SWE English session."

pi-setup

Configures Pi Agent (@earendil-works/pi-coding-agent) with DeepSeek V4 (built-in provider) and Ant-Ling Ring-2.6-1T (custom OpenAI-compatible provider), plus a curated extension set, while automatically working around 7 known pitfalls.

It detects before it acts: it probes whether pi is installed, whether config already exists, and where keys live before changing anything. Keys are never written in plaintext — pi config only stores lazy !shell commands and the key stays in the user's original file. Model selection, default model, default thinking tier, and the extension set are all confirmed interactively.

Install:

npx skills add https://github.com/wquguru/skills --skill pi-setup

Trigger it in Claude Code by saying "配置 pi", "setup pi agent", "把 ring/deepseek 配到 pi", or run /pi-setup manually.

Prerequisites:

  • Node.js / npm (to install pi)
  • A DeepSeek key (in ~/.deepseek or an environment variable)
  • An Ant-Ling/Ling key (LING_API_KEY, in a shell rc or environment variable) — only needed when enabling Ring

free-my-cpu

Safely audits CPU, load average, I/O wait, and Docker container pressure on Linux hosts, especially production-like servers over SSH. It follows the same read-only-first posture as free-my-disk: collect repeated samples, separate CPU-bound work from I/O pressure, inspect hot containers and logs, then classify possible fixes by impact before changing anything.

Install:

npx skills add https://github.com/wquguru/skills --skill free-my-cpu

Ask for help with prompts like:

  • "Use free-my-cpu to inspect this remote Docker host."
  • "Why is this Docker host load average so high?"
  • "Find which container is burning CPU."
  • "Use this Grafana URL and the last 6 hours to check whether monitoring queries are causing load."

youtube-toolkit

Downloads YouTube videos with yt-dlp and post-processes them with ffmpeg: fetch best-quality streams, convert containers (e.g. AV1/webm → H.264 mp4), extract audio, and burn in translated subtitles with a translator watermark.

It is macOS-first and bakes in the gotchas learned the hard way — chiefly that Homebrew's lean ffmpeg ships without libass/drawtext (so subtitle burn-in needs ffmpeg-full), that YouTube auto-captions are overlapping rolling cues that must be re-segmented before they read cleanly, and that styling via an ASS file avoids the force_style comma-parsing trap. Two helper scripts do the deterministic work: install-tools.sh and burn-subs.sh.

Install:

npx skills add https://github.com/wquguru/skills --skill youtube-toolkit

Ask for help with prompts like:

  • "Download this YouTube video to ~/Downloads."
  • "Convert it to mp4."
  • "Add Chinese subtitles and a @handle 翻译 watermark."

Prerequisites:

  • Homebrew (macOS) or apt/pip (Linux)
  • yt-dlp and a libass-enabled ffmpeg — install-tools.sh sets both up

Local Development

For local hacking, the Justfile is self-contained: it symlinks every skill under skills/ and external/ into Claude Code (~/.claude/skills/), Codex (~/.codex/skills/), and Grok (~/.grok/skills/), so edits take effect immediately without re-publishing. No private sibling repo is required.

just add      # symlink all skills (skills/ + external/) into agent dirs
just remove   # remove only the symlinks that point back into this repo
just status   # show install status for each skill in all destinations
just vendor   # pull third-party skills declared in external.yml into external/
just          # list available recipes

add is idempotent and never clobbers: an existing real directory or a symlink from another install tree is left untouched and reported as skipped (foreign symlink). remove only removes symlinks that point back into this repo. After adding a new skill folder, re-run just add.

External (third-party) skills

skills/ holds the skills authored here (and published via npx skills add). Third-party skills from other repos are not copied into git — they are vendored separately so this public repo stays purely self-written and free of foreign licenses:

  1. Declare sources in external.yml — a repo, a ref (branch/tag/commit), and the sub-paths to pull. Each dir's basename, prefixed with 3rd-, becomes the skill name (e.g. dashboarding → 3rd-dashboarding).
  2. just vendor shallow/sparse-clones each source, copies the named skill dirs into external/ (git-ignored) under their prefixed names, rewrites each SKILL.md frontmatter name to match, and records the exact upstream commit SHA in external.lock.
  3. just add symlinks them alongside the self-written skills.

Only external.yml + external.lock are committed; the vendored code is not. Re-run just vendor any time to refresh (or after pinning a ref). A name that would collide with a self-written skills/<name> is skipped; entries removed from external.yml are pruned from external/ on the next just vendor.

Repository Structure

Justfile          # add / remove / status / vendor (self-contained)
external.yml      # third-party skill sources (repo + ref + sub-paths)
external.lock     # resolved upstream commit SHAs (auto-generated)
scripts/
  vendor.sh       # pulls external.yml sources into external/
skills/           # self-written skills (published)
  claude-5-best-practice/
  gpt-5-6-best-practice/
  english-swe-daily/
  …
external/         # vendored third-party skills (git-ignored; just vendor)
index.html        # public site

License

Self-authored skills under skills/ are intended for public use via this repo and npx skills add. Third-party code under external/ is not redistributed in git — only external.yml / external.lock are tracked; run just vendor to fetch upstream under their own licenses.