Workshop

01 Open-Source Skills & Tools

Most of these are agent skills — small, sharply-scoped capability packages that teach a general-purpose AI agent to do one thing properly. Each one started as a tool I built to scratch my own itch, then open-sourced. A few have found a real audience on GitHub.

• wrap-skill · ★65 — An archive skill for AI coding sessions, built around one idea: keep a single living document instead of letting docs drift out of sync with the code. It folds scattered, stale markdown into one CLAUDE.md, reconciles it against the actual codebase, and pins a PROGRESS block tracking project state — so documentation and execution stop telling different stories, and every session ends cleanly archived with one source of truth to resume from.
• promptly-prompt · ★52 — A remake of my earlier prompt-optimization project, now a hook that injects a discipline before each turn. For anything complex it makes the AI restate the requirement in its own words first — because the 99% of prompting that actually matters is just clarity and consistency, and the fastest route to both is to have the model voice its understanding and align before acting. Once aligned, you can push it a step further to spell out its own logic and reasoning.
• researcher-skill · ★29 — A deep-research skill built on one spine: a vertical axis (depth-first) tracing how something became what it is, and a horizontal axis (breadth-first) sweeping it against its peers — crossed into a single judgment. It's built to resist hype, weak summaries, and one-sided narratives, with each round of search following the leads the last one surfaced rather than fanning out shallow.
• xmum-room-booking · ★6 — An agent skill that auto-books the university library's study rooms: it logs in, reads the captcha with Gemini, and reserves a slot across all four room types. Pair it with a scheduler and the AI runs the booking at a fixed time every day, hands-free.
• xmum-application-skills · ★1 — A suite of Codex skills for graduate-application planning — background diagnosis, school-list strategy, statements, CVs, recommendations, region comparison — orchestrated by a main advisor skill that tracks state and routes each question to the right specialist.
• ai-era-style-cards · ★1 — AI-era style cards: 140 poster-design styles laid out as comparison cards. Born from being fed up with the generic posters and covers AI spits out — instead of gambling on a prompt, you browse 140 styles side by side, pick the aesthetic you actually want, then generate. A style vocabulary for GPT Image / Midjourney / Canva.
• andrej-karpathy-skills — A fork of the Karpathy-inspired CLAUDE.md (five principles countering common LLM coding pitfalls — silent assumptions, overcomplication, careless edits). I added one rule I live by: always reuse before you build. Identify which domain the problem belongs to and pull in the field's existing knowledge and solutions first — because without that domain knowledge you can't write the right prompt in the first place; reusing is how you acquire it.