What is a8-loom-coordinator?

a8-loom-coordinator is a set of configuration files, rules, and automated checks that turns Claude, or another capable AI assistant, into something closer to a senior engineer working on a real software project, while a human stays in charge, approving decisions and spot checking the work rather than watching every step. It works with any programming language or framework, on either the frontend or backend of an application. The project grew out of problems the author ran into while using an AI assistant on a shipping application over several months. AI assistants sometimes edit code from memory rather than the current file on disk, so this stack blocks an edit until the assistant has actually looked at the real, current version of the relevant files. It also limits how much background information loads into the assistant's context at once, since loading everything buries the one rule that matters, by keeping a small always loaded core and pulling in extra instructions only when relevant keywords appear. It also addresses AI generated reviews that confidently report false findings, by requiring any claim to be checked against the actual code before it is trusted. A central idea is matching each AI model's cost to the difficulty of the task: the most capable, expensive model handles judgment heavy work like reviewing changes and making architectural decisions, a strong model handles deep investigation and building features, a mid tier model handles repetitive, rule based sweeps across many files, and the cheapest model handles simple lookups. If a cheaper model fails, the work escalates one tier rather than jumping straight to the most expensive model. The stack is organized into folders covering governance templates, reusable skills, automated hooks that enforce the rules regardless of which AI model is in use, tools that index existing code so nothing gets duplicated, and an optional module for adding features to an existing frontend without breaking its design. Installation is available as a Claude Code plugin, as an npx command that scaffolds a configuration file into any repository, or by cloning the repository and running an install script by hand. The project is released under the MIT license.