Show HN: A working reference implementation of context engineering

It breaks down context engineering into five components—design, retrieval, injection, output, and enforcement—and demonstrates it as a reference implementation that runs end-to-end with an actual corpus.

• Corpus: Organizes materials such as the organization's ADRs, code, and standards into a corpus.
• Retrieval: Finds document fragments relevant to the request.
• Injection: Puts the found context into the model's working memory.
• Output: Produces reviewable artifacts like code, PRs, and documentation.
• Enforcement: Verifies that the generated output actually reflects the context that was referenced.

This repository uses the Spring PetClinic codebase and ADRs as examples, using Claude and Titan on Amazon Bedrock. Each folder has its own requirements.txt and README.md, and ready-to-run commands are also provided.

Running it requires Python 3.11+, AWS credentials, and a region that supports Claude and Titan. It notes that Anthropic Claude requires a one-time First Time Use (FTU) procedure per AWS account, and that an AccessDeniedException may occur on first run.

The repo also explains the difference between RAG and context engineering. It states that having only steps 1 through 3 is RAG, while adding output and enforcement on top of that makes it a reviewable and governable context engineering.