Agent skills — drop-in skills for coding agents
Open-source skills for Claude Code, Codex, Cursor, Gemini CLI and other agents. Install one, or all of them with one command.
Overview
In plain terms: a skill is a folder with one instruction file, SKILL.md, that your AI agent reads when the task fits.
They work with Claude Code, Codex, Cursor, Gemini CLI, Copilot, OpenCode, and chat apps that take skill uploads. They all live in one public repo, MIT licensed.
| Skill | What it does | Say |
|---|---|---|
| upsweep | Finds Upwork jobs that pass your filters, then sums up what those clients keep asking for. Uses Upwork's official MCP and your own login. | "upsweep" |
| vidpipe | Turns an article into a silent 720p explainer video of at most 60 seconds, paced so every word can be read. | "turn this article into a video" |
| clean-ram | Finds the browsers and dev servers that agent sessions left running on a small Linux machine, and stops only the ones you approve. | "clean ram" |
| close-session | Saves what the next session needs into your memory or notes, compactly, then lists what is still open. | "close the session" |
Install
Any agent (it asks which agents to install into):
npx skills add dave8172/skills # pick from the list
npx skills add dave8172/skills --skill vidpipe # just one
npx skills add dave8172/skills --skill '*' # all of themClaude Code plugin:
/plugin install all --marketplace dave8172/skillsClaude on the web or desktop: download the repo as a ZIP, then upload one skill's folder under Customize → Skills.
upsweep
Say "upsweep". On the first run it reads your Upwork profile and proposes searches and filters in one message; reply "ok" or change anything in plain words. You get the jobs that pass, newest first, with the link and the Connects cost, and a summary of what those clients ask for: the problems, the tools they name, what they screen on, and how fast posts fill up. It reads Upwork only through the official MCP, so there is no scraping. More on its own page.
vidpipe
Give it a URL, a Markdown file or a README. It writes a scene list from the article's own facts and renders a 720p MP4 with animated captions, numbers, charts and diagrams. Every piece of text stays on screen for at least 0.7 s plus one second per three words; if a scene would flash by, the renderer refuses and names it. --check reports the timing in about a second, before any render. Needs Node 18+, ffmpeg and a Chromium.
clean-ram
Agents start headless browsers, dev servers and mock servers, and often leave them running. On a small VPS they pile up until the machine swaps. The skill lists the leftovers with why each looks safe to stop, or why it might still be in use. It stops only what you approve, checks each process ID first, and never touches services, containers, databases or its own process tree. Linux only.
close-session
Before you quit, it finds your memory or notes system (built-in agent memory, CLAUDE.md, AGENTS.md, a notes folder) and saves only what the next session needs: where the work stands, decisions and why, your corrections. One line per fact, under fixed headings. It updates old entries instead of adding duplicates, never stores secret values, and gives you a one-line prompt to resume with.
← All work