skills

My public agent skills for Claude Code and Codex. Each skill is one folder under skills/, and its SKILL.md says when it fires and what it does. Every skill belongs to exactly one category in catalog.yaml. Nothing here holds personal data: guard, lint and CI run before every push.

Install

One skill: link or copy its folder into a repo. Nothing depends on the rest of the tree.

git clone https://github.com/ong6/skills.git
ln -s "$PWD/skills/skills/web-extract" /path/to/repo/.claude/skills/web-extract   # Claude Code
ln -s "$PWD/skills/skills/web-extract" /path/to/repo/.agents/skills/web-extract   # Codex

Everything, per machine: clone into a root folder, then let bin/skills link what the machine's profile enables and keep it current.

git clone https://github.com/ong6/skills.git <root>/skills
<root>/skills/bin/skills up --repo <repo> --machines <machines.json>

As a Claude Code plugin: /plugin marketplace add ong6/skills, then /plugin install skills@ong6-skills.

bin/skills

A single-file, stdlib-only Python CLI (3.7+). It never starts a model session.

Command	Does
up --repo R --machines M [--hook claude|codex] [--wait]	match the profile; clone missing checkouts and run the pull flow in the background; link; register Claude local-scope MCP servers; print one status line (plus reloadSkills hook output when links changed)
link / unlink --repo R --machines M	create the enabled links and remove stale ones / remove every link into the checkouts
doctor --repo R --machines M [--json]	report profile, links, dangling links, dirty checkouts, shelf, usage, last fetch; exit 1 on problems
autosync --repo R --machines M	for a Stop hook: publish every dirty or unpushed checkout with a generated message; exit 2 when the agent must fix a guard hit, lint failure or conflict
publish --checkout C -m MSG	lock, guard (public checkout), lint changed skills, commit, rebase if needed, push
guard [PATH...]	public-safety scan; exit 2 on hits
find [WORDS...] [--json]	search core, rarely used and retired skills and the catalog's external links; ranked, with tier and path or URL
usage --repo R --machines M [--scan] [--json]	last use of every skill across machines and what tidy would do
tidy --repo R --machines M [--dry-run]	shelve idle core skills and restore used shelved ones, then push
shelve / restore NAME...	move skills between skills/ and rarely-used/ by hand, then push

With --machines omitted, the registry is $SKILLS_MACHINES, else life/tech/machines.json or machines.json inside --repo or the working directory's repo (which then becomes the --repo). Commands that link, clone or push (up, link, unlink, autosync, tidy, shelve, restore) exit 2 when neither a registry nor a repo is found, instead of acting on the default profile.

The profile is the skills object of the first machine in the registry whose match fields all equal this host's (os, model, hostname prefix, wsl, user):

{"machines": [{"match": {"os": "Linux", "hostname": "devbox"}, "name": "Dev box",
  "skills": {"scope": "project", "root": "~/src", "private": false,
             "repos": ["~/src/notes"], "enable": ["*"], "disable": ["category:design"],
             "mcp": []}}]}

• scope: project links into <repo>/.claude/skills and <repo>/.agents/skills for --repo and each of repos, and removes our links from ~/.claude/skills and ~/.agents/skills. scope: user links into those two home folders instead.
• root holds the checkouts <root>/skills and <root>/skills-private (default ~/Sideproject).
• enable and disable take skill names, globs and category:<key>.
• readonly: true is for a machine that can read the checkouts but never push them (an agent user on a read-only deploy key): up pulls and links, records no usage, never tidies, and resets a clean checkout to origin, dropping any local commit, so it keeps fast-forwarding; autosync does nothing.
• No matching machine means project scope, --repo only, public skills only.
• Only symlinks that point into the two checkouts are ever created or removed. Anything else in a link folder is left alone and reported as a conflict.
• A Claude Code SessionStart hook can run bin/skills up --repo "$CLAUDE_PROJECT_DIR" --machines <file> --hook claude. up never prompts and always exits 0, and it returns in well under a second when nothing changed.
• A Stop hook can run bin/skills autosync --repo <repo> --machines <file> so skill edits commit and push themselves at the end of each turn. The public checkout is guarded and linted first; nothing is pushed past a guard hit.

Sync and tiers

Pull flow. At most every five minutes up starts a background sync. For each checkout it fetches, then fast-forwards a clean tree, or rebases its local commits onto origin, and pushes commits an earlier auto-sync, tidy or usage scan left behind. Unpushed public commits are guarded again before they leave. A dirty tree is left to autosync, which commits it first.

Tiers. Each checkout keeps its skills in two folders:

Folder	Linked	Holds
skills/<name>	yes	the core set
rarely-used/<name>	as a stub	skills nobody used for 30 days; still callable by name, found by find

Once a day the background sync records which skills this machine's Claude Code and Codex transcripts used, in usage/<machine>.json in the private checkout. It then runs tidy. A core skill moves to rarely-used/ when no machine used it in 30 days, it arrived in core more than 30 days ago, and there is a month of usage data. Pinned skills (pinned: in catalog.yaml) and skills that a staying skill names in backticks stay put. A shelved skill used again on any machine moves back.

A shelved skill keeps a generated stub in the link folders, so typing /name in Claude Code or $name in Codex still runs it. The Claude stub sets disable-model-invocation: true and the Codex stub sets allow_implicit_invocation: false, so neither model loads or picks it on its own. The stub tells the agent to read the real SKILL.md, and that read counts as a use.

A use is a Skill tool call, a /name or $name request, or reading or running the skill's own files. Edits, searches, and sessions that touch more than five skills (audits, evals) are not uses.

Public and private

• Public (this repo): shareable skills with no personal information.
• Private (skills-private, a separate private repo): skills that name their owner's paths, accounts or voice, plus mcp/servers.json and guard-terms.txt. Private skills may compose public ones by name. Public skills never reference private ones.
• When both repos have a skill of the same name, the private one wins on machines that link private.
• bin/skills guard blocks home-directory paths, personal email addresses, phone numbers, NRIC/FIN numbers, private keys, tokens, the owner terms in skills-private/guard-terms.txt and AI-CLI launches (codex exec, claude -p) from skill code. To keep a deliberate line, add public-guard: allow to it.
• Skills are location-independent. They refer to other skills by name and never resolve their own symlink to find siblings.

Included skills

The skill index and catalog below are generated from catalog.yaml and each skill's frontmatter. CI fails when they drift.

Skill	What I use it for
web-extract	Web search and page reads, spending Firecrawl credits first.
youtube-transcript	Fetch clean captions from a YouTube video.
save-video	File a YouTube video as a searchable note with takeaways, category and tags.
datastore	Keep structured records as append-only JSONL with a local SQLite cache for SQL.
refine (rarely used)	Review and improve the work until its constraints and requested score pass.
write-a-brief	Write or tighten a one-to-two-page brief for a meeting with a professional, quotes included.
markdown-to-pdf	Turn a Markdown note into a checked, print-ready A4 PDF.
system-diagram	Turn a real system flow into a polished, readable SVG figure.
ux-design	Audit or design a page's UX against a usability checklist before visual polish.

Catalog

Research

Getting text and data out of the web and video.

Skill	Does
web-extract	Search the web and read or extract pages, spending Firecrawl's monthly credits first and switching to the host's native search and a local readability fetch while they are at zero. Use for web searches, reading a URL, JavaScript-heavy or Cloudflare pages, public PDFs, structured extraction or page interaction. Not for YouTube (youtube-transcript) or LinkedIn (off-limits).
youtube-transcript	Fetch clean transcripts or captions from YouTube URLs or video IDs. Use when the user shares a YouTube link or asks to transcribe, summarize, cite, or read a video.
save-video	Save a YouTube video into the video library of the host repo as a note with overview, key takeaways, category and tags, indexed for later search. Use for "save this video", "add this to my videos", a YouTube link with "keep/remember/save", or "what videos do I have on X" (search the library). Not for a one-off summary the owner only wants to read now (youtube-transcript alone).
datastore	Keep many same-shaped rows as a git-backed dataset: append-only JSONL as truth, a local SQLite cache for SQL. Use for "track X over time", "store these as a table", "query my jobs/prices/workouts", daily ingest scripts or durable machine state. Not for: prose notes, small Markdown tables, secrets.

Elsewhere:

• firecrawl/firecrawl-claude-plugin — Firecrawl's official plugin. web-extract wraps the same CLI.
• tavily-ai/skills — Search, extract, crawl and deep research over the Tavily API.
• exa-labs/exa-mcp-server — Neural web search and page fetch as an MCP server.

Writing

How replies, briefs and documents read.

Skill	Does
refine (rarely used)	Explicit-only. Improve the latest deliverable (draft, code, plan, note, design) through fresh independent reviewer agents until it meets the requested score and all explicit constraints, or three refinement rounds pass. Use for "/refine", "refine this", "run the refine loop", "auto-improve this". Not for a single quick edit, a PR or diff review (code-review), or feedback about agent behaviour (feedback-loop).
write-a-brief	Write or tighten a one-to-two-page meeting brief for a professional (contractor, consultant, doctor, vendor), including what to ask about their quote. Use for "meeting brief", "prep doc", "questions to ask", "check this quote", or fixing a brief that is too long or AI-sounding; not for email drafts or general notes.
markdown-to-pdf	Convert a markdown note into a verified, print-ready A4 PDF. Use for printable notes, meeting handouts, or rebuilding generated PDFs; not for editing existing PDFs or building a LaTeX document.

Elsewhere:

• JuliusBrussee/caveman — Compressed "caveman" prose for agents; cuts output tokens by about two thirds while keeping code and commands exact.
• ayghri/i-have-adhd — Stops the agent burying the answer.
• anthropics/skills — Anthropic's docx, pdf, pptx and xlsx skills live here (source-available, not open source).

Design

Diagrams, interfaces, 3D and visual judgment.

Skill	Does
system-diagram	Draw a polished system or architecture diagram as hand-authored inline SVG: request flows, agent pipelines, queues, service maps, before/after. Use for "system diagram", "architecture diagram", "draw the flow", a figure for a site, README or note. Not for: data charts (dataviz), UI mockups, one-hop relationships.
ux-design	Audit or design a page or app's UX against a usability checklist (navigation, states, touch targets, focus, contrast, motion, mobile), then hand visuals to frontend-design. Use for "UX pass", "audit the UX", "make this usable on mobile", "accessibility check". Not for: aesthetics alone (frontend-design), charts (dataviz).
3d-design	Choose and run a 3D workflow (Blender, Three.js, CSS, model-viewer and others) for illustrations, models, character animation and interactive web scenes. Use for "3D design", "model this", "Blender or Three.js", "seamless loop", "make the movement natural". Not for: page layout, 2D diagrams, Blender install issues.
blender-authoring	Create, edit, animate, export and render Blender scenes using headless Python (bpy). Use after 3d-design selects Blender, or when the user explicitly requests Blender or a blend file. Covers geometry, materials, rigging, animation, lighting and export. Not for choosing between 3D tools, Three.js playback or ordinary page layout.

Elsewhere:

• tt-a1i/archify — Turns a codebase or a description into interactive architecture and sequence diagrams as self-contained HTML.
• nextlevelbuilder/ui-ux-pro-max-skill — UI and UX guidance for agents building interfaces. Large; its 119-rule usability list is the useful part.
• Leonxlnx/taste-skill — Aesthetic judgment for generated frontends.
• emilkowalski-skills by Emil Kowalski — Collection of design-engineering skills for UI polish and motion: building, reviewing and auditing animations, Apple-style interaction, and an animation glossary. Seen in Insane Claude Design Skills You Need To Actually Build Beautiful Sites.
• emil-design-eng by Emil Kowalski — Encodes Emil Kowalski's approach to UI polish, component design and animation decisions; a starting skill when building a new interface. Seen in Insane Claude Design Skills You Need To Actually Build Beautiful Sites.
• animate by Emil Kowalski — Builds one animation in a fixed order of decisions (whether to animate, purpose, tool, properties, curve, duration, interruption, exit) and writes the code; use it to add motion to an existing site or app. Seen in Insane Claude Design Skills You Need To Actually Build Beautiful Sites.
• animation-vocabulary by Emil Kowalski — Glossary that turns a vague description of a motion effect into its exact name, so you can prompt an agent or designer precisely. Seen in Insane Claude Design Skills You Need To Actually Build Beautiful Sites.

Life

Trips, shopping, property, learning and hard choices.

Skill	Does
decide	Make a hard, consequential choice and commit: reframe it, find the unlisted option, rough numbers, downside, pre-mortem, one pick, the fact that flips it and a tripwire. Use for "should I X or Y", "help me decide", "is it worth quitting / switching", job offers. Not for: what to buy (product-shopping), trivial or factual questions.
trip-planner	Plan or replan trips using live research for routes, hours, parking, food, tolls, and pacing. Use for itineraries and on-the-road questions about stops, parking, or eating.
product-shopping	Recommend what product to buy, find the best legitimate price for a chosen model, compare its landed cost across countries (SG vs US and others, in SGD), or say where to buy it. Use for "which X should I get", "is X worth it", cheapest price, deal checks, price history, "should I buy this overseas", "where can I buy X". Not for: property (evaluate-condo).
evaluate-condo (rarely used)	Evaluate a Singapore condo listing or development and file a cited Buy/Neutral/Avoid note. Use for PropertyGuru, 99.co, EdgeProp, "evaluate this condo", or "is this unit worth it".
teach	Create a stateful, multi-session course in the learning workspace. Use only when the user invokes teach or asks to start structured learning of a topic. Explicit-only; not for one-off explanations or research reports.

Career

Interview drilling and job-search practice.

Skill	Does
interview-prep	Interview practice from the candidate's own resume and target-company notes: mocks, STAR answers, system-design walkthroughs, timed DSA reps by due pattern, debriefs of real interviews. Use for "mock interview", "give me a rep", "what's due", "I just had an interview". Not for: live rounds (interview-live), resumes (cv), job hunting.

Elsewhere:

• kirilxd/swe-interview-coach — Behavioural and system-design prep for Claude Code.

Field engineering

Customer-facing decks for forward-deployed work.

Skill	Does
create-fde-deck	Draft or review a customer-facing field-engineering deck: discovery, architecture or pilot readout, every claim labelled evidence, assumption or proposal. Use for "turn these notes into slides", "pilot readout", "go/no-go deck", "review this deck", "fact-check my slides". Not for: interview practice.

Skill building

Making, testing and improving the skills themselves.

Skill	Does
skill-search	Find whether a skill exists for a job, including unloaded ones: rarely used, retired (private archive) and external skills in the catalog. Use for "do we have a skill for...", "is there a skill like X", or a skill named but not loaded. Not for: writing or revising a skill (skillsmith).
skillsmith	Make or revise an agent skill grounded in the current repo, then prove it beats the same model without it: gate, inventory, draft, lint, blinded baseline-vs-skill eval, keep or retire. Use to add, write, revise, merge, review or prove a skill. Not for: typo-only edits, AGENTS.md rules, hooks.
feedback-loop	Log the owner's feedback on how the agent, a skill, hook, rule or reply behaved in feedback.md, and patch its target in the same session. Fires on "don't do X", "stop doing", "next time", "I prefer", "that was wrong", "always"/"never" about the agent. Not for: feedback on the owner's writing or a one-off instruction.

Elsewhere:

• anthropics/skills — The official reference set and the Agent Skills spec.
• mattpocock/skills — Matt Pocock's engineering set. grill-me, tdd, to-spec, implement, code-review, diagnosing-bugs.
• obra/superpowers — Brainstorm, plan, TDD, review as one methodology.
• dietrichgebert/ponytail — Pushes the agent to reuse the standard library and platform features before adding code or dependencies.
• addyosmani/agent-skills — Production-grade engineering discipline for coding agents.
• multica-ai/andrej-karpathy-skills — One CLAUDE.md distilled from Karpathy's notes on LLM coding pitfalls.
• ong6/groundplane — My deterministic-boundary library for agent output. Not a skill, but the reference for what an agent may generate versus what code must produce.
• ComposioHQ/awesome-claude-skills — Curated index of Claude skills and tooling.
• skill-creator by Anthropic — Creates new skills, improves existing ones, and runs evals to measure a skill's performance and how reliably its description triggers. Seen in The NEW Agentic OS standard for Claude 5 Models is here (Full Breakdown).
• cli-printing-press by Matt Van Horn — Generates agent-first command-line tools for an app or API, driven by its /printing-press skills, so an agent gets a connector without an MCP server. Seen in The NEW Agentic OS standard for Claude 5 Models is here (Full Breakdown).

Licence

MIT for everything here. Linked repos carry their own licences. Contributions are welcome; see CONTRIBUTING.md. Report security issues privately as described in SECURITY.md.