Should you use cwc-workshops?
Use cwc-workshops for Claude agent exercises when you can take responsibility for maintaining copied examples. Anthropic's cwc-workshops materials are not maintained and do not accept contributions. The repository README makes that policy explicit, so a successful installation should not decide whether you adopt the code.
The collection includes separate exercises for model selection, agent decomposition, Claude Code product workflows, managed agents, memory, evaluation, and research applications. Choose a folder for the design question you want to study. Our result for agent-battle/ does not establish that the rest of the collection runs.
What did our sandbox establish?
Our run on September 27, 2026 recorded a 16-second install for agent-battle/ at commit 068b84b. The full recorded commit is 068b84bb03d2ae87c51edb2837dda25c84c1d686. Our sandbox used 3 CPUs, 8 GB of memory, and the Python 3.12 Bookworm uv image. Its stored summary records 67 dependency packages occupying 79 MB and a build step lasting 4 seconds.
Our harness records command outcomes through success status, elapsed time, exit code, timeout status, and available log tails. Installation and build each returned exit code 0 without timing out. The install tail identifies the Python environment; the build tail is empty. Those outcomes establish that the recorded commands completed, but do not demonstrate a working game, CLI deployment, MCP connection, or managed-agent session.
The harness skipped tests with the reason "no script". That is neither a failed test nor proof that the repository contains no tests. These stored parser summaries do not establish exhaustive unique test coverage. We performed no new execution or complete-log reconciliation for this correction.
Pip-audit recorded 0 advisories, but the supplied record does not establish which packages were scanned. Scan coverage is unknown. The dependency count cannot fill that gap, and a zero advisory total is not a security assessment.
The MrKeyoor testing-methodology page explains how command outcomes and their limits reach our reviews. Its supplied October 2026 description is provenance context, not evidence that the September run used that procedure version.
Which exercise should you inspect first?
Start with agent-decomposition if your question is how to divide instructions, computation, and delegation. Its workshop guide starts with a 402-line inventory-agent prompt and provides setup instructions, evaluation tasks, an editing loop, and session transcripts for debugging. The practical value is that participants check behavior after changing the design. The guide's reported improvements are upstream workshop results, not results from our sandbox.
For generated interfaces, inspect how-we-claude-code. The root README describes a Vite and React application whose components emit a machine-readable DOM contract for runtime checks by an agent or CI. Agent Battle addresses a different problem: configuring a managed agent that drives a local game bot over MCP. Our command outcomes did not verify either workshop's runtime behavior. Before borrowing a pattern, reproduce the relevant workflow and check its current prerequisites.
Should you choose courses or recipes instead?
Anthropic Courses recommends a learning order beginning with API fundamentals. Its README then lists prompting, prompt evaluations, and tool use. Choose that sequence when you need foundations before a workshop; it can also complement cwc-workshops when an exercise assumes knowledge you have not learned yet.
Claude Cookbooks provides code and guides intended for integration into your own projects. Its README requires a Claude API key and says the examples are primarily Python. Choose it for an application recipe rather than a conference exercise. It is a complementary reference, not validation of the workshop code or a demonstrated production replacement.
What do the scores mean?
Community scores 1 because the repository explicitly declines maintenance and contributions. Documentation scores 4 because the supplied decomposition guide explains setup, iteration, evaluation, and debugging. Setup ease scores 3: our commands completed in one folder, while the collection contains separate projects requiring their own preparation. Maturity scores 2 because the evidence establishes teaching material without verified application workflows from our run.
These are editorial assessments, not benchmark results. The rubric is 1 for a major limitation, 2 for substantial constraints, 3 for usable with preparation, 4 for strong supporting evidence, and 5 for comprehensive evidence in the assessed dimension. Setup considers preparation and observed command outcomes; documentation considers exercise guidance; community considers the participation policy; maturity considers demonstrated suitability beyond teaching.