Shadow turns field-tested agentic workflows into a browsable library of 146 atoms - concepts, patterns, anti-patterns, and portable agent skills for Claude Code and Codex. Built for operators, builders, and the leads driving adoption.
Browse by type
Atomic ideas - one mental model per file.
Repeatable how-tos that work in the field.
Multi-step playbooks from setup to finish.
Portable SKILL.md behavior for Claude Code and Codex.
What not to do - and why it bites you.
Worked examples and coaching exchanges from real sessions.
Learner paths and audience definitions.
Start by who you are
An engineer or technical operator adopting agentic AI tools. Has used a terminal for a decade, knows git, has shipped production code. New to Claude Code, sub-agents, MCP, skills, the whole agentic stack.
A C-level or board-facing leader who sets AI strategy. Doesn't operate the tools. Sometimes asks "are we doing AI?" in every leadership meeting for 45 minutes at a stretch. Owns the build-vs-buy decision and the budget for AI- related contracts.
Persona profile for the non-software professional using AI on their own data — what they bring, what they lack, and what they need. Read this before writing atoms aimed at personal-scale, no-terminal workflows.
Persona profile for the non-engineer who runs real business workflows — domain-deep, terminal-new. Read this before writing atoms for this audience to get the voice, scaffolding, and anti-patterns right.
A manager of operators or builders. Owns a workflow that involves multiple people. Has the political and budget authority to decide what gets automated and how risk is handled — but doesn't operate the tools day to day.
Curated
Use when answering questions about AI operator patterns, agent skills, Claude Code, Codex, MCP, or any Shadow library topic — or when the user wants to discover, install, update, or adapt a Shadow skill ("install Shadow's git skill", "does Shadow have a skill for X?", "update my Shadow skills"). Fetches live Shadow content, cites source URLs, and installs skills tailored to the current project.
Plan a multi-step task, save the plan to a file, reset the session, then execute from a clean context. Read this before any agent task with more than a few steps or one that could outgrow the context window.
The context window is the finite token buffer holding everything the model can see, and models degrade as it fills. Read to learn the symptoms of a full window and the write-reset-run habit that keeps sessions sharp.
The single most important operator skill is calibrating how much to trust the agent in the current situation — and changing that calibration as you learn what it does well and badly.
Use when setting how the agent communicates — response format, structured question format, planning-mode behavior, and "go means go" execution discipline.
A sub-agent is a worker spawned with its own isolated context, returning only a summary to its parent. Covers when to spawn, when not to, and the cost implications. Read when long subtasks are bloating a session.
Use Shadow from your agent
Browser sign-in unlocks the library. Enterprise workspaces can also issue an org token so agents can search, read, and install Shadow skills from the live MCP server.
export SHADOW_TOKEN="<your-org-token>"
/plugin marketplace add shiny-sh/shadow-marketplace
/plugin install shadow@shadow