skills

Public agent skills, one folder each under skills/<name>/ (SKILL.md, a README.md for people, plus any scripts). This file is public, so keep it publication-safe: no personal data, no private project names, no machine-specific paths.

• Install paths that must keep working: npx skills add dave8172/skills [--skill <name>] (scans skills/*/SKILL.md) and the Claude Code marketplace in .claude-plugin/marketplace.json (one plugin per skill plus all). A new skill gets a row in README.md, a plugin entry for itself, and a path in all.
• upsweep: Upwork's API & MCP Terms are the hard boundary. It runs on demand only, uses the user's own criteria with no ranking of its own, stores no job data, only reads during a sweep, and submits a proposal only after the user approves Upwork's preview, one at a time. Any change must keep all of that. The clauses are cited in its SKILL.md. Test it like a new user: use an empty folder with .claude/skills/upsweep symlinked to skills/upsweep, a fresh UPSWEEP_DIR, and every Upwork write tool in --disallowedTools. Say "upsweep", then "ok". Afterwards, check that nothing except profile.md was written.
• vidpipe: the limits are the product. Keep it 720p, silent and at most 60 s, and keep the reading rule (0.7 s + words ÷ 3) and the size floor (MIN_FS). The renderer refuses specs that break them. Never add a bypass flag. Test: in skills/vidpipe/app, run npm install, then node render.mjs ../examples/dns.json --check, then --preview, then a full render. A new scene type must declare all its text in plan. Demo mode (app/demo.mjs, DEMO.md): real runs only (answers come from record, never hand-written), the same reading rule on captions and cards, one-line captions, explain then act (a caption only over a still scene; MOVES flags one over an action), at most 180 s; no bypass flags. Several pages join only as whole films with a cut between them; never cut inside a film. Test: record, film, compose --preview and a full render of examples/demo/; after a change to scrolling or capture, also film a page that scrolls (no blank-frame warning).
• clean-ram: generic Linux only. Don't name particular services, hosts or users. Nothing is stopped without the user's approval of a numbered list; other agent sessions may still use an orphaned process.
• close-session: stays agnostic about where memory lives. The user's own system always wins; memory-system.md is only a fallback, read on demand. Keep SKILL.md short, since it loads in full whenever the skill is used. It saves and reports; it never stops processes or deletes files (that is clean-ram's job, with approval).
• Libraries take version ranges, never pins. Skills stand alone and point to nothing outside this repo.