# Start here Try a small AI context workflow: capture a decision, ask a bounded question, check the answer and keep a reusable handoff. A no-install walkthrough for keeping useful context. Write a handoff with a goal, decision, source, uncertainty, next step and constraint. Ask an agent one bounded question about a named repository file. Check each citation and command against the source. Record a test result separately from a command that has only been documented. Keep a small continuation handoff and a separate detailed log. Read [State That Survives the Session](/blog/state-that-survives-the-session/) for the architecture and its measurement gaps. Inspect the [256-run experiment](/blog/what-docs-llms-files-and-indexes-actually-save/) before assuming that more files or indexes save money.