You ask the agent a simple question — how does this thing work? — and what comes back is technically correct, three paragraphs long, and slides off your brain like water off a window. You read it twice. You still couldn't repeat it to a colleague.
The agent isn't wrong. It's just not communicating.
That gap — between knowing a topic and being able to say it like a human would — is what these skills are for. Small, opinionated instruction files that change how the agent talks: one analogy not three, the answer before the journey, short sentences, no "at its core…" filler. The agent already knows the material. The skill just gets it to land.
Portable across Claude Code, Codex, Cursor, Gemini CLI, and any other
Agent Skills-compatible tool. Each skill is
a folder with a SKILL.md file, loaded only when relevant — so you
can keep many on hand without bloating context.
Three families. Pick the one that matches what your reader needs.
- eli5 — for when you want to get it. Plain language, one strong analogy. Triggers on "ELI5", "in simple terms", "plain English", "dumb it down".
- exec-summary — for when a busy leader has 60 seconds. Leads with the decision, not the mechanism. Triggers on "exec summary", "TL;DR for my boss", "for leadership", "one-pager".
- eli-engineer — for when a peer engineer wants precision over accessibility. Names the data structure, the invariant, the trade-offs. No analogies. Triggers on "ELI-engineer", "technical explanation", "skip the analogy".
- bluf — Bottom Line Up Front. The answer in the first sentence, supporting detail after. Triggers on "BLUF", "TL;DR", "lead with the answer", "get to the point".
- minto-pyramid — for proposals, recommendations, and decision docs. Answer first, then grouped reasons, then detail. Triggers on "structure this proposal", "Minto", "pyramid principle", "make this argument tighter".
- feynman — for checking your own understanding. Explain it plainly, then identify the part where you got fuzzy. Triggers on "do I actually understand this", "Feynman", "test my understanding", "what am I missing".
- diagram-it — for when a relationship is the point and prose buries it. Picks the right diagram kind (flow, sequence, hierarchy, state), keeps it under five nodes, emits mermaid. Triggers on "draw this", "diagram", "visualise", "show me the flow", "picture this".
- compare-table — for when the reader is choosing between 2–4 options and prose buries the differences. Four columns max, five-word cells, takeaway above, nuance below. Triggers on "compare", "what's the difference", "X vs Y", "should I use A or B", "side by side".
- slide-deck — for when you're presenting to a room. One idea per slide, headline = takeaway (not topic), speaker notes carry the story, no filler slides. Triggers on "make slides", "deck", "presentation", "slidify this", "turn this into slides".
More coming, slowly.
Install as a plugin to get every skill at once and pick up updates automatically:
/plugin marketplace add patrick204nqh/skills
/plugin install patrick204nqh@skills
Copy the skill folder you want into one of these locations:
~/.claude/skills/ # Personal (Claude Code)
.claude/skills/ # Project-local, shared via git
~/.agents/skills/ # Codex / generic Agent Skills
Then just talk to your agent normally — skills trigger on intent, not slash commands (though most can be invoked explicitly too).
The best explanation is the one a colleague can repeat an hour later. The best summary is the one a leader can act on without scrolling. The best diagram is the one you don't have to squint at. These skills are opinionated rules — one analogy not three, the answer before the journey, five nodes not fifteen — because vague guidance doesn't change behaviour. Specific guidance does.
Each skill stays in the same shape: a single instruction file that shapes how the agent communicates — in words, in structure, or in a picture. No tools, no rendering, no dependencies. The skill tells the agent what to produce; the host renders it. Match the medium to what the reader needs.
Built by Patrick. Steal them, fork them, send a PR if you sharpen one.
MIT