dotfiles

Run the codex ↔ developer review cycle on a GitHub PR autonomously — fetch codex findings on the latest commit, evaluate each fairly on the merits, fix legitimate ones with project-appropriate commits + tests, push, post a fresh `@codex review` top-level PR comment, and either schedule the next wakeup (continue) or return without scheduling (stop). Designed to be invoked via `/loop /codex-review-loop` so the `/loop` harness provides the cross-invocation cadence; the skill itself does exactly one cycle per call and decides whether there is a next cycle. Use whenever the user says "codexのレビュー対応して", "codex とのラリー回して", "codex review loop", "@codexと回して", "codexのレビュー対応ループ", or otherwise indicates they want the codex review rally handled end-to-end. Also use proactively right after the user creates a PR that codex will review, if they signal they want the back-and-forth handled. Handles codex's 👀-reaction heartbeat: a 👀 on your re-request comment means codex is processing; absence after a reasonable wait means the me

Creates a pull request by analyzing the diff between the current branch and origin/main (or master). If a PR template exists, follows it. Asks for language preference first, presents a draft for approval, and only pushes/creates PR after user confirmation. Use when the user says "create PR", "make a pull request", "submit PR", or similar.

Generates a visual, step-by-step explanation of any concept, code, architecture, or process the user asks about. Saves the result as a Markdown file in a tmp directory so the user can view it in a Markdown previewer with rendered Mermaid diagrams. Use this skill whenever the user says "/explain", asks you to explain something in detail, wants a visual walkthrough, or requests a diagram-rich explanation. Also trigger when the user asks things like "how does X work?", "walk me through Y", "explain Z step by step", or "I don't understand W" — especially when the topic is complex enough to benefit from diagrams.

Generate raster images from text prompts using Google's Gemini / Imagen API (Nano Banana = gemini-3-pro-image, and Imagen 4). Use this whenever the user asks to create, generate, draw, make, or design an image, illustration, picture, icon, logo, hero/cover/banner art, background, texture, avatar, concept art, or any visual asset — including phrasings like "画像作って", "イラスト描いて", "図を作って", "アイキャッチ", "サムネ", "挿絵", "ビジュアル". Also use PROACTIVELY (without being asked) when building a website, slide deck, document, README, or learning material that would clearly benefit from illustrations or visual assets. Supports reference images for editing and for keeping a consistent art style across many images. Note: for precise charts/graphs/diagrams with exact text or math, prefer SVG or code — AI raster generation garbles text and equations.

LAPRAS MCPを起点に、外部ソースの横断調査・本人へのヒアリング・デザイン方向の提案・実装までを一気通貫で行い、その人にしかない「ナラティブ・ポートフォリオ」（キャリアを物語として体験させる自作Webページ）を作り上げる。ユーザーが「narrative-portfolio」「ナラティブポートフォリオを作りたい」「LAPRASのデータで自分のポートフォリオ/年表/キャリアの物語を作って」と言ったら必ず使う。LAPRAS MCPが設定されている環境で、ポートフォリオ・経歴の可視化・自己紹介ページ・キャリアの振り返りコンテンツの相談を受けたときも、明示的な指名がなくてもこのスキルを提案する。

Turn a feature idea, bug, refactor, or UI tweak into a zero-ambiguity GitHub Issue that a junior-level worker AI (Cursor's Composer 2, or a similarly cheap model) can implement without making a single meaningful decision. Grills the user relentlessly one question at a time, researches the codebase, compares libraries, and records rejected alternatives so the worker can't drift mid-implementation. Use this skill whenever the user asks to "ticket this", "make an issue", "spec this out", "hand this off to Composer / Cursor / the worker", "チケット化して", "Issueにして", "prepare a spec", or gives you a GitHub Issue URL and asks to update the description. Make sure to use this skill for ANY handoff from planning (Opus 4.6) to implementation (cheap worker model) — features, bugs, refactors, or UI — even when the user doesn't explicitly say "issue" or "ticket" but the intent is to write a worker-ready spec.

Build a multi-page, research-grade VitePress learning site from scratch for a topic the user wants to deeply understand. Drives an interview-first workflow — grills the user one decision at a time on scope, audience, language, depth, and boundaries; gathers primary sources before any writing begins; scaffolds a VitePress project with Mermaid (and optionally KaTeX); dispatches parallel writing sub-agents under strict fact-first discipline (mandatory source citations per file, explicit uncertainty markers, zero speculation); runs an independent external audit pass (e.g., via a second LLM/agent such as Codex) for fact-check and structural critique; verifies rendering in a real browser with dev tools (build success alone is insufficient — math layout, Mermaid SVGs, console errors all get checked). Use this skill whenever the user asks to "build a learning site", "make a study reference", "produce a structured explainer for X", wants a VitePress-based "I want to deeply understand X" outcome, or says things like "ド

Execute whitelisted commands on the host machine from inside a container