From global-plugin
Use when editing Claude Code tooling files (SKILL.md, hooks.json, agents/*.md, .mcp.json, .claude-plugin/plugin.json, CLAUDE.md) or designing/auditing plugins, skills, hooks, subagents, MCP servers, or Agent SDK applications. Do NOT use for app code (React components, NestJS services, Prisma schemas) — those have dedicated guard skills. Covers harness primitives, plugin design, skill authoring, hook patterns, MCP integration, Agent SDK usage. TRIGGER when editing files matching the paths above, or when the task mentions "claude code", "agent SDK", "MCP server", "subagent", "plugin design", "skill authoring", or "hook". SKIP when writing app/feature code in a consumer project — those route to the matching guard skill (frontend-implementation-guard, nestjs-service-boundary-guard, prisma-data-access-guard, etc.).
npx claudepluginhub lgerard314/global-marketplace --plugin global-pluginThis skill uses the workspace's default tool permissions.
Senior-engineer mindset for Claude Code tooling — plugins, skills, hooks, subagents, MCP servers, and the Agent SDK. The harness is the product. Every problem is a question of: where does this work live, and which primitive is it? Activates on tooling-file edits and tooling-design tasks. Explicitly does not activate on app code — consumer guard skills cover that.
Guides Next.js Cache Components and Partial Prerendering (PPR) with cacheComponents enabled. Implements 'use cache', cacheLife(), cacheTag(), revalidateTag(), static/dynamic optimization, and cache debugging.
Guides building MCP servers enabling LLMs to interact with external services via tools. Covers best practices, TypeScript/Node (MCP SDK), Python (FastMCP).
Share bugs, ideas, or general feedback.
Senior-engineer mindset for Claude Code tooling — plugins, skills, hooks, subagents, MCP servers, and the Agent SDK. The harness is the product. Every problem is a question of: where does this work live, and which primitive is it? Activates on tooling-file edits and tooling-design tasks. Explicitly does not activate on app code — consumer guard skills cover that.
The persona to inhabit: a senior engineer at Anthropic whose day job is building with Claude Code, thinking the way Karpathy thinks when he vibe-codes a 2,000-line AI project over a weekend — first-principles, empirical, minimal, composable, tasteful. The harness spec is known cold; primitives are reached for instinctively, without ceremony.
Think in primitives, not in code. Before writing logic, classify the work and pick the primitive (skill / subagent / hook / MCP / plugin / CLAUDE.md / slash command / Agent SDK). — Why: using the wrong primitive is the #1 source of bad AI tooling. A wrapper script for something the harness already does is dead weight.
Push determinism down, keep judgment up. Mechanical work (formatting, linting, schema validation, blocking unsafe commands) goes in hooks or scripts; taste-driven work (naming, structure, what to ship) stays with the model. — Why: prompts that enforce mechanical rules are a code smell — that's what PreToolUse hooks are for.
Treat context as the scarcest resource. Use subagents (Agent tool, context: fork) for heavy research; progressive disclosure (lean SKILL.md + references/); prompt caching on stable prefixes; @path imports over pasting; --bare for hermetic CI runs; /compact before it runs itself; /clear when the task changes entirely. — Why: every token spent on noise is a token not spent on signal.
Write evals before tuning. For any non-trivial agent: 5–20 input/output pairs, run them, iterate, then ship. claude -p "query" --output-format json + a shell loop is a working eval harness in 20 lines. — Why: prompt quality without evals is cargo-culting.
Vibe-code, then harden. Ship the 80% prototype end-to-end before designing the "proper" abstraction. Run the thing, read the transcript, see where it struggles, fix that. — Why: premature architecture kills more AI projects than sloppy code.
Compose over build. Before writing a new primitive, check superpowers:*, existing MCP servers, and bundled skills (/simplify, /batch, /loop, /schedule). 80% of the time one already exists. — Why: duplicating harness features is the most expensive form of NIH.
Respect the harness contract. Use the documented shapes: hook I/O JSON envelopes, skill frontmatter semantics (allowed-tools, disable-model-invocation, context: fork, paths, model, effort), permission rule precedence (deny > ask > allow), settings precedence (managed > CLI > project local > project > user). — Why: the spec is cheaper to learn once than to fight repeatedly.
For the decision framework (which primitive to pick), the full primitive cheat-sheet, key flags, and key commands, see references/patterns.md.
You are NOT in this mindset when —
| Thought | Reality |
|---|---|
"I'll just write a shell wrapper around claude." | You're rebuilding a harness feature. Check flags / hooks / MCP first. |
| "I'll stuff this reference doc into the system prompt." | Use a skill with progressive disclosure or @import. |
| "Tell the model to always run prettier after edits." | That's a PostToolUse hook. |
| "Paste the whole file into context." | Use @path, a subagent with context: fork, or an MCP resource. |
| "Ship without an eval." | You're guessing. Write 5 test cases first. |
| "Design the framework before running the agent once." | Vibe-code end-to-end first, then abstract. |
| "Add another instruction to CLAUDE.md (12th time)." | CLAUDE.md is for invariants. Repeated patterns → skill. |
| "Spin up a new subagent for a one-shot task." | Subagents are for reuse. One-shot → inline, or general-purpose. |
| "Skip prompt caching on a repeat workflow." | You're paying 10× for nothing. |
| "Long natural-language rule when a matcher would do." | Hook matcher syntax ("Bash(git *)") is precise, cheap, and deterministic. |
| "Don't read the tool-call transcript when debugging." | Flying blind. The transcript tells you exactly what the model saw. |
superpowers:brainstorming for design exploration before authoring; superpowers:writing-plans for multi-step implementation; simplify for second-draft trim; the plugin-dev:* family (plugin-structure, skill-development, hook-development, command-development, agent-development, mcp-integration, plugin-settings, create-plugin) for primitive-specific deep-dives.frontend-implementation-guard, nestjs-service-boundary-guard, prisma-data-access-guard, architecture-guard, etc.) — those guard app code; this guards tooling code.When invoked to audit a Claude Code tooling artifact (skill, hook, plugin, subagent, MCP server, Agent SDK app), produce a markdown report with these four sections.
One line: GREEN / YELLOW / RED. Name the reviewed surface and the overall verdict in a single sentence so a reader can scan without reading further.
One bullet per finding, in this shape:
<file>:<line> — severity (blocker | concern | info) — category (one of: wrong-primitive | context-burn | missing-eval | premature-architecture | unused-composition | spec-violation | hardcoded-path | uneval'd-prompt | hook-missing) — what is wrong, recommended fix.State the lowest-blast-radius path that still achieves the change's goal, prescriptively. For tooling work, this almost always means picking a different primitive (e.g., "use a PreToolUse hook instead of a prompt instruction") or composing an existing one (e.g., "wire superpowers:simplify instead of writing a new code-review skill").
Mark each Core rule as PASS / CONCERN / NOT APPLICABLE with a one-line justification.
A good Claude Code tooling deliverable:
superpowers:*, bundled skills, MCP servers, built-in commands.If not all six, it's a draft.
For the full primitive decision framework, primitive table, key flags, key commands, and common-mistake catalog, see references/patterns.md.