skills

Hand-rolled Skills, and other AI bits I build from my own day-to-day experience.

Token estimates use tiktoken's o200k_base encoding. SKILL.md is the entry prompt; the total adds every other Markdown file an agent can go on to read. Scripts, assets and config ship with a skill but are run rather than read, so they are left out.

Skill	SKILL.md	Total
afk	228	228
cicd-fy	1,517	15,838
coding	1,048	1,048
fast-feedback-mode	394	394
git	2,414	14,814
grill-me	61	61
grilling	578	578
kubernetes-debug	240	1,513
mise-fy	2,524	36,888
review	325	325
skill-fy	2,901	6,389
variants/afk-careful	89	89
variants/afk-soon	166	166
variants/afk-yolo	114	114
variants/watch-pr	660	660
variants/watch-pr-ci	261	261
variants/watch-pr-comments	256	256
vendor/domain-modeling	781	1,939
vendor/gh-stack	2,282	7,946
vendor/handoff	198	198
vendor/humanizer	7,056	9,335
vendor/i-have-adhd	1,743	1,743
vendor/improve-codebase-architecture	1,349	2,945
vendor/prototype	642	3,546
vendor/show-me	783	783
vendor/teach	1,997	3,855
vendor/wait-what	102	102
vendor/writing-for-agents	2,349	2,916

Skills

1. coding | My coding and docs conventions and taste. Loaded first on anything that touches code or docs.
2. mise-fy | Encode mise best practices. And Transform projects into using mise + hk for a good dev setup!
3. cicd-fy | CI/CD patterns: build/test/scan, label-driven releases, changelogs from PRs, and publishing signed + attested artifacts.
4. git | My git conventions and PR workflow.
   • watch-pr | Babysit a PR: fix CI and answer reviews, until green and quiet (or forever).
   • watch-pr-ci | CI only: fix what goes red, stop at green.
   • watch-pr-comments | Reviews only: answer them as they land.
5. kubernetes-debug | WIP: Failure mode when debugging k8s
6. afk | Letting the agent know I am no longer gonna be there to answer questions (un-grill-me lol).
   • afk-careful | Conservative with its autonomy.
   • afk-yolo | Max non-destructive autonomy.
   • afk-soon | I am leaving soon, ask all ur questions now...

Vendored

Skills I didn't write but use daily, vendored under skills/vendor/ so the plugin ships one set instead of asking you to install them separately. Thanks to mattpocock/skills, humanlayer/skills, blader/humanizer, and github/gh-stack.

skills-lock.json is the list, and mise run skills:sync rebuilds the directory from it. It records where each skill comes from, not which version, so a sync takes whatever upstream has now.

Models

I mostly use (opus-4.8|5,gpt 5.5,gpt 5.6 sol|terra, and grok 4.5), weaker models might not work as good, especially that I rely on model's world knowledge over explicit examples.

Installation

npx skills for just the skills. The plugin to get my whole setup.

npx skills (recommended)

Installs skills into any SKILL.md-aware agent. You don't need a plugin host.

npx skills add sherifabdlnaby/skills              # all skills
npx skills add sherifabdlnaby/skills --skill git  # just one skill
npx skills list sherifabdlnaby/skills             # see what's available

Claude Code (plugin)

/plugin marketplace add sherifabdlnaby/skills
/plugin install skills@sherif-plugins

Update with /plugin marketplace update sherif-plugins; manage from /plugin.

Cursor (plugin)

/plugin marketplace add sherifabdlnaby/skills
/plugin install skills

Manage from /plugin. Same repo, same skills — Cursor reads .cursor-plugin/.

---

Development

This repo is managed by mise: it pins every tool, exposes the tasks, and wires the git hooks, so you only deal with mise once.

brew install mise                                  # or: apt/dnf/pacman install mise, winget install jdx.mise
echo 'eval "$(mise activate zsh)"' >> ~/.zshrc     # bash: mise activate bash >> ~/.bashrc
mise doctor                                        # confirm the install is healthy

Other install methods: installation docs.

You also need the gh CLI logged in (gh auth status) — mise doesn't install it, and the pinact check step borrows its token to resolve action tags to SHAs. Without it that step runs unauthenticated and gets rate-limited.

One-time setup, from the repo root:

mise trust && mise run setup   # trust config, install tools, self-install the pre-commit hook

Linters, formatters, and validators run via hk:

mise run check          # report on uncommitted changes (alias: mise run lint)
mise run check --fix    # apply fixes
mise run check --all    # whole repo
mise tasks              # discover every task

Commits run the fast commit gates on staged files, a push runs the slower push gates, and CI runs both as mise run check. lychee checks local/relative links only by default (see .config/lychee.toml).

Releases

Claude/Cursor marketplaces install straight from git and key the plugin on the version field in .claude-plugin/plugin.json and .cursor-plugin/plugin.json — there's no published artifact. Cutting a release:

1. Bump both manifests' version to the same value in your PR. mise run version:check (also an hk pre-push gate) fails if shipped skills//hooks/ content changed without the bump.
2. Merge to main. CI notices the version moved past the latest vX.Y.Z tag and cuts the tag plus a GitHub release, with notes generated from merged PRs and grouped by label: skill, hooks, fix, docs, ci, deps.
3. No version bump ships no release. A PR's sticky comment previews which outcome a merge would produce.

There's nothing to verify or download: the marketplaces install the repo at the tag.

Contributing

See CONTRIBUTING.md.