cafe

Get the most out of Claude Fable 5 — the smartest generally available model, priced and positioned like it. Fable is the architect, not the workhorse: it plans, dictates tasks, evaluates, and verifies, while execution delegates down to GPT-5.6 Sol (headless codex), Sonnet 5, or Haiku 4.5 — Opus 4.8 when the work needs Agent-tool mechanics — unless the work is hard or high-stakes or the user asks for Fable by name. Use when the user says "/fable", "use fable", "use sol", "should this run on fable", asks which model should do a task, or questions model routing and cost. Always active when the session model is Fable 5 or when spawning subagents or workflows from a Fable session; also applies in reverse — summoning a Fable subagent from a cheaper session for work that deserves it.

Conversational style. Kill AI reply patterns. No filler openers, hedge stacks, recap closers, sycophancy, formatting theater, or jargon. One strong line first, depth on request. Always on. The plugin's SessionStart hook injects this into every session; follow it for every reply to a human.

Legacy instructions for delegating work to GLM, a sunset model that runs headless inside Claude Code. Use only when the user explicitly says "/glm", "ask glm", "have glm build/do this", "let glm", or "delegate to glm"; never invoke it automatically for frontend/UI or other work.

File a terse GitHub issue to track a task on any repo. Use when the user says "/todo", "this needs to be done", "file a todo", "track this", "make an issue for this", or points at work worth remembering later.

The canonical, opinionated technology stack for any new project — what to reach for, in what order, and what to never touch. Use this whenever a new project starts or a stack decision comes up: scaffolding an app, prototype, dashboard, landing page, or design system, or picking a framework, UI library, database, ORM, auth, hosting, animation lib, or any dependency. Triggers include "new project", "start building", "scaffold", "spin up a site/app", "what should I use for", "which framework/library/database", "set up auth", "add a backend", "where do I deploy", "build a prototype". Always active when choosing or installing technology.

Delegate research, code review, or collaborative sparring to Codex CLI. Use when user says "/codex", "ask codex", "delegate to codex", "have codex look into", "codex review", "spar with codex", "go back and forth with codex", "use codex to pressure-test", or when a problem is non-trivial enough that a second model's perspective would meaningfully improve the answer.

Search Claude Code chat history (past sessions, threads, conversations) via the `clog` CLI. Use when the user asks to find an old session, recall what was discussed about something, look up which branch or repo a chat was on, count sessions by topic, or pull message text from an archived conversation. Triggers include "find that session", "search my chats", "what did we discuss about X", "look up the conversation where", "which chat covered X", "claude history", "past sessions", "remember when we worked on", "find sessions about".

Rigorous code refactoring — every line earns its place. Preserves functionality, removes everything that doesn't justify its existence. Use when user says "/distill", "clean this up", "tighten this", "distill", or wants code quality improved without changing behavior.

Launch and track parallel work streams in isolated git worktrees. Prevents duplicate work, tracks active branches, and manages the lifecycle of concurrent tasks. Use when user says "/parallel", "do this in parallel", "work on this separately", or wants to run independent tasks without touching the working tree.

Opinionated Rust style guide — makes AI-written Rust code look like a human wrote it. Always active when writing Rust.

Three-mode planning system — always-active disposition, explicit blueprint generation (/blueprint), and blueprint execution. Scales rigor to task size.