Supervise a difficult Codex task through a separate visible worker task. Use when the user asks for a supervisor/worker workflow, delegated implementation with oversight, evidence-based steering, independent verification, or distinct goals across Codex tasks.
Use when Claude should ask OpenAI's Codex CLI for a focused second opinion, independent codebase analysis, architecture review, bug-hypothesis check, implementation critique, or planning help from a separate model. Defaults to read-only headless consultation via `codex exec --sandbox read-only`.
Write prompts, system instructions, agent directives, slash commands, and skill descriptions using two stacked layers — outcome-first (define the destination, success criteria, stopping condition) plus directional language (every sentence names the path with positive verbs). Use when writing or reviewing any prompt, system message, AGENTS.md, CLAUDE.md, skill description, SKILL.md body, tool description, slash command body, eval rubric, or anywhere an LLM reads instructions. Triggers on "write a prompt", "improve this prompt", "audit this system prompt", "outcome-first", "success criteria", "directional", "make this prompt positive", or when authoring any new skill, agent, or directive. Keywords: prompt engineering, system prompt, outcome-first, directional, positive verbs, stopping condition, prompt audit.
Answer questions about how the pi coding agent itself works by reading pi's bundled offline docs and citing the file. Use when asked how pi does X, how to configure pi (settings.json, models.json, providers, keybindings), what a pi flag/mode/feature does, or where a pi feature is documented. Keywords pi docs, how does pi, configure pi, pi settings, models.json, pi providers, pi extensions, pi sessions, pi modes, pi skills.
Author Agent-Skills-standard SKILL.md files designed to be executed by the pi coding agent — often a small local model (llama.cpp/Qwen). Covers pi skill anatomy and discovery, writing low-freedom procedures with machine-checkable validators for small executors, the author-pilot-refine distillation loop driving pi headlessly, and optional Codex co-authoring. Use when writing or refining a skill that pi will run, or when asked to make a skill work well on a local model. Keywords pi skill, SKILL.md, pi coding agent, local model, llama.cpp, qwen, skill distillation, /skill, pilot, refine.
Use the pi coding agent CLI (@earendil-works/pi-coding-agent) and navigate its docs and source. Covers interactive/print/json/rpc modes, model & tool flags, sessions, custom models.json (local llama.cpp/Ollama/vLLM), skills, extensions, and the pi-mono monorepo layout. Use when running `pi`, driving it headlessly/programmatically, configuring local models, writing pi extensions/skills, or answering "how does pi do X". Keywords pi, pi.dev, pi coding agent, earendil, pi-mono, llamacpp, models.json, rpc mode, json mode, qwen.
Reference for Claude Code features and best practices: writing CLAUDE.md, skills, hooks, sub-agents, agent teams, headless mode, scheduled tasks, worktrees, batch, frontend verification. Use when authoring/auditing CLAUDE.md, configuring Claude Code, writing skills, or orchestrating agents. Keywords CLAUDE.md, skill, hook, sub-agent, agent team, headless, worktree, /loop.
Design, build, review, or debug article and web demos across the full range from static to live: semantic HTML/CSS layouts, responsive components, design systems, Grid/Flexbox, accessibility AND interactive Canvas 2D, SVG, WebGL/Three.js, physics/mechanical simulations, sliders, scrubbers, requestAnimationFrame loops, React/MDX/Deno/TanStack embeds. Use when building a non-interactive figure or page layout, a live model-driven explainer (Bartosz Ciechanowski style), or anything between. Keywords: demo, graphics, canvas, WebGL, Three.js, animation, simulation, HTML, CSS, layout, Grid, Flexbox, semantic markup, responsive, accessibility, interactive explainer.