Claude Code changes the dynamics of AI-assisted coding by addressing a longstanding problem.
For more than often, developers experience their sessions end before every requirement has actually been completed. One of the latest additions tackling this issue is the goal command, which lets developers define a precise, verifiable endpoint for a task.
Instead of relying on the model to decide when work "feels done," Claude continuously evaluates progress against the stated condition after every turn and keeps iterating until the requirement is satisfied.
The idea builds on the community-driven “Ralph loop” approach, but Anthropic has now integrated it directly into Claude Code so no custom scripting is needed.
How do you keep Claude working until the job is done? Claude Code helps with this in a few ways, including one we shipped recently: /goal. pic.twitter.com/QtVPmwoKct
— ClaudeDevs (@ClaudeDevs) May 13, 2026
The syntax is intentionally simple. A developer can enter something like:
goal all tests in test/auth pass and lint is clean.
Once set, the goal becomes a persistent checkpoint for the session. A lightweight evaluator, which can be powered by faster model such as Haiku, reviews the transcript after each step.
If the goal has not been met, Claude explains why in a short status message and automatically continues working without waiting for additional prompts. When the condition is finally satisfied, the goal clears itself and the session records a compact “goal achieved” summary including time, turns, and token usage.
It's the Ralph loop, built into Claude Code. Every time Claude tries to stop, it checks your condition against the transcript.
Not done? It keeps going. Done? You get a "Goal achieved" summary.— ClaudeDevs (@ClaudeDevs) May 13, 2026
Users can inspect the active goal at any time with `goal`, or stop it using commands like goal clear, stop, reset, or cancel.
The system is intentionally constrained to information visible inside the transcript, meaning goals must rely on explicit outputs such as test results, command logs, lint checks, or file changes rather than hidden external state.
Conditions are capped at 4,000 characters to keep evaluations lightweight and reliable.
/schedule kicks off Claude on a cadence. Nightly test runs, morning triage, weekly cleanup.
— ClaudeDevs (@ClaudeDevs) May 13, 2026
What makes the feature especially powerful is how well it integrates with Claude Code's broader automation toolkit.
Auto mode reduces approval interruptions during long sessions, stop hooks allow external validation like CI checks, loop enables repeated prompts for iterative tasks, and schedule supports recurring runs such as nightly reviews. Combined together, these features enable extended unattended coding sessions with clearly defined stopping conditions.
The result is a workflow that feels far more persistent and outcome-oriented than traditional AI coding interactions.
Developers can define measurable objectives like “every call site compiles and tests pass” or “all migration tasks are complete,” while still maintaining boundaries and transparency throughout the process.
Long-running only works if Claude doesn't have to wait on you. That's why we have auto mode, which you can enable with shift + tab in CLI or by using the mode selector on desktop.
Docs here: https://t.co/g9R3v5f8UJ— ClaudeDevs (@ClaudeDevs) May 13, 2026
More importantly, this highlights how quickly is shipping capabilities for Claude Code.
The pace of feature releases has become remarkably fast, with the platform evolving from a simple coding assistant into something much closer to an autonomous development environment.
Features like goal may end up being a genuine gamechanger for AI-assisted software engineering because they shift the interaction model from one-off prompting toward persistent, self-correcting execution focused on verifiable outcomes.