context-engineering: Build Governed AI Workflows on Amazon Bedrock

Overview

OutcomeOps’ context-engineering repository is a Python reference implementation built on Amazon Bedrock, using Anthropic Claude for generation and Amazon Titan for embeddings. It demonstrates how to give AI systems relevant organizational knowledge, such as architecture decision records (ADRs), so their outputs reflect local standards rather than generic patterns.

The project frames context engineering as more than retrieval: it connects a corpus and retrieval pipeline to context injection, reviewable outputs, and enforcement. The examples use Spring PetClinic and ADRs, making this a practical reference for teams evaluating how to make AI-generated software work more consistent with their decisions.

Key Features

• Implements a five-part model: corpus, retrieval, injection, output, and enforcement.
• Uses ADRs and a Spring PetClinic codebase as a shared example corpus.
• Runs examples with Anthropic Claude and Amazon Titan through Amazon Bedrock.
• Includes comparisons of context-engineered output with output produced without context, and discusses differences from RAG and other tools.
• Provides runnable examples organized by component, each with its own README and requirements file.
• Includes case studies and an example of an A/B comparison where ADRs change generated code.

Use Cases

• An engineering team can prototype ways to supply architecture decisions to an AI coding workflow before applying the pattern to its own corpus.
• A platform or AI engineering team can examine how retrieval, model input, output review, and enforcement fit together beyond a basic RAG pipeline.
• Technical leaders evaluating context-engineering claims can use the examples to inspect the approach and compare it with alternatives.
• Teams already using Amazon Bedrock can adapt the examples to test whether organizational standards improve generated work.

Getting Started

The examples require Python 3.11 or newer and an AWS account with credentials configured. From the repository root:

cd 01-corpus
pip install -r requirements.txt
python ingest_adrs.py ./sample-adrs

See the README and the component folders for setup details and additional commands.

Alternatives

• ace: ACE builds reusable agent playbooks from task outcomes, while context-engineering retrieves organizational decisions and checks generated work against them.

Considerations

• Examples depend on Amazon Bedrock, AWS credentials, and a supported region. Anthropic Claude also requires a one-time First Time Use form for an AWS account that has not used Anthropic models on Bedrock.
• The project is marked quiet: it has no commits in the last 90 days and no release in over a year. It has one recorded contributor, so adopters should assess maintenance risk and verify behavior before relying on it.
• The supplied project facts report no CI or tests. The repository is presented as a working reference implementation, not as a packaged service with those checks reported.
• The MIT license permits broad reuse, subject to its license terms.