Show HN: How context engineering works, a runnable reference

What happened

A new open-source reference implementation of "context engineering" has landed on GitHub, and it aims to stop chatty AIs from guessing your playbook. The repo walks through a full pipeline — corpus, retrieval, injection, output, enforcement — using a Spring PetClinic example and Amazon Bedrock models. Short version: treat context as a first-class engineering artifact, not as a prompt you type when you remember.

The repo and the idea

The codebase is the companion to an essay on outcomeops.ai and is built to show the pattern end-to-end with Anthropic Claude for generation and Amazon Titan for embeddings. All examples run on Bedrock; each folder implements one of the five components (01-corpus through 05-enforcement) and includes its own requirements and a runnable command. It has been reported that teams that adopt this model often discover the harder problem isn't technical plumbing but changing roles, KPIs, and decision rights — because if your AI can read the corpus, org dynamics shift.

Running it (quickly)

You can clone the repo, ingest sample ADRs, and run examples using Python 3.11+ and an AWS account that supports Claude/Titan. There are a few Bedrock quirks to know: Anthropic Claude requires a one-time First Time Use form and, allegedly, filling it in via the Bedrock catalog grants immediate access with no review queue. The README walks through prerequisites, region choices, and a quickstart so you can see retrieval + injection shaped outputs for yourself — and compare the same task with and without context engineering.

Why it matters

Why care? Because generic assistants still hand you answers that need local tweaking; context engineering promises outputs that already conform to your standards and decisions, and an enforcement layer that makes those dependencies auditable. Sounds boring? Maybe. Useful? Absolutely. In an era obsessed with RAG, Copilot, and agent frameworks, this repo is less flash and more blueprint — practical code you can run to see what changes when context stops being a back-of-envelope afterthought.

Sources: github.com/outcomeops, Hacker News