Anthropic Builders

What's in this pack

This is the first-party Anthropic stack — every asset here was authored or sanctioned by Anthropic itself. If you want to know "what does Anthropic actually ship to help me build with Claude?", this pack is the answer.

This pack distinguishes itself from boris-cherny-files (a personal workflow) by being policy-level: it's the bundle Anthropic itself recommends as a starting point for new Claude developers.

Why install the official bundle

Three reasons. First, trust — these assets are kept current by Anthropic and don't drift. Community subagents go stale; Anthropic-authored ones are updated alongside the model. Second, coverage — the 17 skills cover the common nouns (code, docs, web, search, repo, file, image, slide, sheet…). You rarely need to write the basic skill yourself. Third, idiom — the SDK shows the canonical loop and the CLAUDE.md template encodes the conventions Claude Code expects. Reading them is the fastest way to write idiomatic agent code.

Install in one command

# Pull the entire pack
tokrepo install pack/anthropic-builders

# Or grab pieces
tokrepo install claude-agent-sdk-typescript
tokrepo install claude-md-template

The skills land under .claude/skills/, the Agent SDK is added to your dependencies as a normal npm / pip package, and the CLAUDE.md template is dropped at the project root for you to fill in. Quickstarts ship as separate top-level directories (quickstarts/customer-support/ etc.).

Common pitfalls

• Confusing skills with subagents. Skills are folders with a SKILL.md plus optional scripts; subagents are markdown agent definitions invoked with @name. Both are valid; this pack ships skills, not subagents.
• Outdated SDK assumptions. The Agent SDK was renamed from "Claude SDK" — old tutorials reference the old name. Trust the GitHub README over any tutorial you find on Medium.
• CLAUDE.md as bible. The template is a scaffold. Customize it heavily for your repo; otherwise Claude Code will follow generic guidance instead of project-specific rules.
• Quickstart-as-product. The five quickstarts demonstrate patterns; they are not production-ready apps. Treat them like example code, not boilerplate.
• MCP spec creep. MCP evolves; check the spec version printed by the SDK and match the version of clients you target. A version skew between server and client surfaces as silent tool-call failures.

Relationship to other packs

• Boris Cherny Files — personal flow built on top of this bundle.
• MCP Server Stack — concrete servers that speak the protocol referenced here.
• Prompt Engineering Toolkit — pairs naturally for prompt-quality work inside agents built with this SDK.

If you only have time to install one pack from TokRepo, this is a strong default — it sets the foundation everything else assumes.

When this pack alone isn't enough

The official bundle is canon, but it is intentionally generic. Three needs it does not address: domain skills (legal review, medical coding, finance — write your own or grab a community pack), production observability (the SDK's loop is open-loop by default; you need a tracer), and team-shared prompt libraries (CLAUDE.md is per-repo, not per-team). Layer the relevant TokRepo packs over this base and you have an end-to-end stack instead of a starter kit.

A note on the 17 skills

Anthropic does not publish a numbered roster — the count of 17 reflects the assets currently shipped through their docs and example repos. The exact list rotates as new ones land and older ones merge. This pack tracks the live set: when Anthropic ships a new official skill, the pack manifest updates and tokrepo upgrade pack/anthropic-builders pulls it in. You don't have to re-discover the roster each release.