sinanonur/longrun

This repository is not a software application but an operating manual for running Claude Code, an AI coding agent, on long and largely unsupervised development sessions. The main artifact is a playbook file that you place into your project. The playbook gives the agent a set of rules and habits so it can keep working productively on tasks for extended periods without losing track of what it has done or what it needs to do next. The approach relies on treating the disk as the agent's memory. The playbook instructs the agent to maintain three files: a working task queue, a frozen list of features and their acceptance criteria, and an append-only log of decisions and progress. By reading and writing to these files, the agent can survive context compaction, which is when earlier parts of a conversation get summarized or dropped. The agent follows a strict per-task loop: implement the work, verify it with tests and linting, commit the change, and log what happened. The rules explicitly forbid skipping verification, silently retrying the same failure, or expanding the scope of work mid-run. The playbook also includes a system for routing work to the right AI model. Instead of always using the most expensive option, it provides a scored table of models based on cost, intelligence, and taste. It defines when to delegate mechanical work to a cheaper model and when to keep judgment calls in the main session. The repository explains how to bridge Claude Code to other model families, specifically GPT from OpenAI and Gemini from Google, using community plugins. This allows bulk work to run on a cheaper model while Claude stays the conductor, verifying and reviewing the results. To use the manual, you clone the repository to a stable location on your computer and import the playbook file into your project's configuration. The author notes that you should edit the playbook before adopting it. The model roster and its scores will go stale as models change, and some file paths in the document are personal references you will need to replace or remove. You also need to create the task, feature, and log files yourself before starting your first session.