tlogben/baransu

Converts any content source into a browser-ready, Kami-themed HTML book with SVG diagrams — Acquire (URL / slug / local path / text) → Synthesize → Render (quality-gated). Trigger On '$baransu:book', '轉成 book', '做成 HTML book', '存成 book'. Not for editable Markdown output ($baransu:read to capture, $baransu:learn to digest) — $baransu:book only emits rendered HTML.

Ports Claude Code skills, plugins, or marketplaces one-way to OpenAI Codex format (SKILL.md / batches / complete plugin content including bundled agent TOMLs). Trigger on 「轉成 codex 版」「給 codex 用」「port to codex」, or questions about Claude→Codex field mapping (disable-model-invocation, context fork, ARGUMENTS, plugin.json). Not for reverse porting (Codex→Claude) or authoring a brand-new Codex skill from scratch that isn't a port of an existing Claude source.

Writes a one-page work contract (~35 lines) before a medium-sized change: goal, assertable criteria, can't-miss surfaces, verbatim constants. Use before implementing a feature that deserves pinned acceptance. Trigger On '$baransu:contract', '寫合約', '一頁合約', '開工合約', 'pin the criteria'. Not for an unsliced multi-module effort (slice it first, then one contract per slice) or post-hoc verification (use $baransu:seal). 繁體中文輸出。

Improves, scores, or evolves a SKILL.md via a forward-only ratchet — fixed 9-dimension rubric, one single-variable change per round, three fresh blind judges; kept only on strict improvement, else the snapshot is restored. Dual-axis evaluation (structure + effectiveness), held-out validation, Kami result card; adoption is an Authorization PAUSE. Not for authoring a brand-new SKILL.md, or deciding whether a skill should exist ($baransu:think 存廢判決 Kill/Keep/Pivot). Trigger on '$baransu:evolve', '優化 skill', 'skill 評分', '演化 skill', 'optimize skill', 'improve skill quality', 'evolve a skill', '幫我改 skill'. 繁體中文輸出。

Audits a project's agent configuration and AI-coding maintainability — instruction drift, hooks/MCP, verifier surfaces, code-rot signals — via a budget-aware five-layer audit; escalates to inspector subagents only for deep audits. Also the DEFAULT DISPATCH for environment problems — a skill's CLI/SDK missing, command not found（環境急診：預設修法＝官方一步全域安裝）. Trigger On '$baransu:health', '健康檢查', '配置體檢', '檢查配置', 'AI 可維護性', 'agents ignoring instructions'. Not for verifying baransu's own skill structure (scripts/verify-skills.py), reviewing a single model output ($baransu:review), or substituting for lint/typecheck. 繁體中文輸出。 Also use when: 檢查 claude, 檢查 codex, 配置對不對, 健康度, AI coding 腐化, 程式碼變爛, 上下文混亂, 驗證缺失, hooks 沒生效, MCP 壞了, AGENTS.md, agent instructions, check config, audit config, health check, config drift, command not found, 環境問題, 依賴缺失, SDK 沒裝

Tracks a bug from symptom to root cause before any fix: picks the right observability tool (playwright / MCP db / LSP / logs / static analysis), bisects, confirms or discards hypotheses before touching code. Trigger On 「排查」「查 bug」「追問題」「為什麼失敗」, 'debug', "what's wrong", 'not working'. Not For: subjective UI taste (→ $baransu:ui); worth-fixing value calls / 值不值得修 (→ $baransu:think 存廢判決 Kill/Keep/Pivot). Also use when: 排查, 查查, 報錯, 崩潰, debug, why broken, not working, fix error, 找 bug, 追問題, 查問題, 狩獵, 定位根因, bisect, 為什麼失敗, what's wrong, hunt the bug

Produces a structured learning brief from any content: a 5-column digest brief per source plus an optional filled outline, from URLs / --topic / captured slugs / mixed. Use when the user wants sources digested into a learning note. Trigger On '$baransu:learn', '研究主題', '整理筆記', '學一下'. Not for capturing raw offline Markdown only (→ $baransu:read --save) nor producing a browser-ready HTML artifact (→ $baransu:book).

Reads any web page, file, or search result into the conversation as clean Markdown; keeps an offline copy under .codex/read/ only with --save or an explicit save request. Use when the user wants a page, PDF, or doc fetched, read, or converted. Trigger On '$baransu:read', '抓網頁', '看一下這篇', '轉成 markdown', '存下來', '存檔'. Not For digesting into notes ($baransu:learn) or browser-ready HTML ($baransu:book).

Independent re-verification of a concrete artifact, diff, or directly quoted claim: dispatches a fresh, narrow-context verifier — with additional perspective agents only for distinct consequential risks — never edits the target, and reports examined scope, findings with evidence, independence, and remaining limits; a clean review is a valid result. Trigger On 「看一下」「看看」「幫我看」「check 一下」「review 一下」, or casual "take a look at X". Not for auditing the user's own project agent-config (route to $baransu:health) or closing a contract-banded task (use $baransu:seal). 繁體中文輸出。

Post-implementation seal, run as a dispatcher: assembles the verification payload, runs the baseline, dispatches a verify-only seal-agent in a clean context (criteria audit / unpinned-surface scan / cross-UI / constants byte-diff / mutation spot-check), fixes findings in the main session with capped re-verification, and stamps the sealed marker on a clean pass. Use to close out a /contract-banded task. Trigger On '$baransu:seal', '封緘', '收尾驗收', '驗收剛做完的', 'seal it'. Not for cross-perspective re-verification of any model output (use $baransu:review) or pre-work criteria pinning (use $baransu:contract). 繁體中文輸出。

Wraps up a session: archives baransu working dirs under .codex/ (except read/learn/book/design products) into the gitignored, local-only .codex/archived/, commits, pushes (optionally $baransu:ship BRANCH), and tears down the worktree once work is on origin. Trigger On '$baransu:ship', '收工', '上傳收尾', '結束這輪'. Not For writing copy ($baransu:write) or reviewing output ($baransu:review) — $baransu:ship only wraps up a session.

Deliberates before building when the user has an undecided idea, feature, refactor, or direction and asks how to design it, which way to go, or whether it is worth doing (想一下, 幫我想, 怎麼設計, 值不值得, 該不該做, 「我想做 X 但還不確定」). Also takes a decision already made when the user wants it tested (拷問我, grill me). Aligns on a restatement of what they actually want, then leaves a stance with its bets named and a plan file under .codex/think/ for review; never code, never a handoff to implementation. Not for debugging an existing error ($baransu:hunt) or pinning acceptance for settled work ($baransu:contract). 繁體中文輸出。

UI/UX design lead for the code in your repo. Use whenever the user wants a page or component to look better, less generic, more designed, or styled after something — 美化、調 UI、加樣式、看起來太 AI／太模板／太陽春、make it look designed, add some styling — it builds or reshapes the web UI with deliberate palette / type / layout / motion choices and writes or edits the UI files directly; learns a design language from a reference (a website, a screenshot, an art style) and refines an existing UI against it. Trigger On '$baransu:ui', '設計 UI', '美化', '調 UI', '介面設計', '照這個網站的風格', 'styling', 'beautifying any web UI', 'frontend design'. Not For: Claude Design 畫布 mockup（內建 design skill）; 說明性圖表與動畫（$baransu:book）; 小寫 design.md 技術架構文件.

Refines existing text, generates a piece from a prompt, or proofreads a document into an error-report HTML (book-styled 錯字修改.html findings table). Auto-classifies input as Refine / Generate / Proofread; follows language prefix or auto-detects. Use for bilingual zh/en writing help. Trigger On '$baransu:write', '潤稿', '寫一篇', '改寫這段', '校對', '找錯字', '抓錯字', 'proofread'. Not for committing finished text ($baransu:ship) or digesting sources into notes ($baransu:learn, $baransu:read).