Warp Agent Working Playbook: From Rules to a Reviewed Diff

In practiceAn AGENTS.md at the repository root, with more specific files in subdirectories. The filename must be in capitals, and WARP.md still works for older projects. /init generates a first draft. The most specific file wins over the root file, and both win over global rules in Warp Drive.

In practiceA named profile per kind of work, each with a base model and a level per action: Agent decides, Always ask, Always allow, or Never. Add read-only commands to the allowlist and keep rm, curl, wget, and eval on the denylist, which forces a prompt even under Always allow.

In practice/plan for any change larger than a small fix. The plan opens in an editor and saves each agent revision as a version. Run one phase at a time. A builder rebuilt an ML project in 18 stories with Warp's agent using specs, AGENTS.md, and repository skills. The builder still had to catch a model input bug.

In practiceRun until completion runs every suggested command until the task ends, and by default it also runs denylisted commands. Turn off the setting that lets auto-approve bypass the denylist, and on Enterprise enforce it from the Admin Panel.

In practiceCodebase Context indexes Git-tracked repositories, each worktree separately, and uses .warpindexingignore to skip files. Free indexes 3 codebases, Build and Max 40, and Business 80. Secret Redaction, off by default, masks matching patterns before a request leaves the machine.

In practiceInteractive code review in Warp, so the agent can act on comments before the change lands. Tests and CI on every change, and a named human reviewer on the pull request. Review comments can also go to Claude Code or another CLI agent running in the same tab.