Context Engineering: Because Your LLM Doesn't Know Your Rules

AI coding assistants write decent code. They just don't know your team's rules. OutcomeOps released a reference implementation of context engineering on GitHub that tackles exactly this. The repo breaks context engineering into five components: Corpus (your org's knowledge), Retrieval (finding relevant pieces), Injection (feeding them to the model), Output (generating reviewable artifacts), and Enforcement (making sure the output actually follows the rules). Components one through three are basically RAG. The last two are what make this different.

It runs on Amazon Bedrock with Anthropic's Claude for generation and Amazon Titan for embeddings. You can clone the repo and run it against a sample corpus of Architecture Decision Records from a Spring PetClinic codebase. The documentation frames context as "a first-class engineering artifact, version-controlled, retrievable, and enforceable, rather than as prompts typed into a chat window."

This idea is gaining traction. Anthropic calls the problem managing a finite "attention budget" to prevent "context rot" in long-context windows. Shopify's Tobi Lütke popularized the term. LangChain's Harrison Chase says providing the right information at the right time is becoming the most critical skill for AI engineers. They're all circling the same idea: prompting as we knew it is over. Managing what goes into the context window is the core engineering problem now.

OutcomeOps is betting this becomes essential for enterprise teams, especially in regulated industries. Feeding an LLM your team's actual decisions at the right moment produces something you can ship using long-term memory.