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.

@bcherny why no bun??

Set a completion condition, and Claude keeps working until it's met: /goal all tests in test/auth pass and the lint step is clean

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.

There’s also /loop to run Claude on repeat. Great for iterative refactors, cleanups, or burning down your backlog.

/schedule kicks off Claude on a cadence. Nightly test runs, morning triage, weekly cleanup.

A stop hook gives programmatic control over when Claude is allowed to finish. Run your test suite, hit a CI endpoint, gate on whatever you want.

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:

If you want to track your "/goal" progress for Claude Code, Codex, OpenCode and Pi etc, I built an open source agent view for that

16 minutes in: Opus 4.7: Well I think we've hit a good stopping point for the day, we should be proud of everything we got done!

@bcherny This is like watching Mark Zuckerberg steal features from literally everyone in 2014. It’s a strategy I guess.

It's literally 24 hours between claiming that it exists and shipping this feature exactly as it was designed in the Codex.

Joker—no one can finish their own Goal because they hit their limit long before the code is even done.

Now just combine that with this powerful auto-permission granting system and you’re all set…

This is what we’ve been waiting for! Finally, Claude didn’t give up easily 😭 /goal: Complete all backlog tasks + 90% test coverage + clean code Thanks, ClaudeDevs team, this is a real game changer for those who like to push Claude until the morning 🙌

@bcherny Well to be honest, I get API rate limit error so I don’t get the job done.

Everyones busy even Clawd.

long-running claude sessions are wayy better when they actually know my values/history/users/taste we're building @heyhyperai to make long-running agents actually work :)

the hard part isnt keeping an agent running its knowing when done is actually done long running agents need a judge that isnt the same model trying to declare victory completion without independent verification is just a fancy hallucination loop

Great tip. The /goal command is basically setting a mission briefing for Claude before it starts. This matters for long tasks: without a clear goal, Claude can wander. With /goal, it self-corrects against the objective. If you are doing multi-file refactors, pair it with /compact to keep context clean.

Goal-seeking is stronger when it includes constraints, not just the finish line. In brownfield code, the hard part is finding hidden dependencies before the agent optimizes for the wrong success condition. Otherwise, “goal achieved” can become technical debt.

I can't use this because I have a $20 plan and my usage limit isn't enough, but with Codex's $20 plan, this feature won't consume my usage limits even if the model I selected is in high mode.

Claude: “these tests don’t seem to working, let’s delete them”

The dangerous part of goal-driven AI isn’t missing the target - it’s hitting the target the wrong way. In brownfield systems, there are usually hidden dependencies and years of assumptions buried in the code. Without clear constraints, “goal achieved” can slowly turn into technical debt.

the mid-task drift was genuinely the most annoying thing lol, testing this today

Now you can give separate tasks to both Claude and Codex at the same time and go do other things, then just come back when they're finished

The underrated bit is giving Claude a finish line it can keep re-reading. Otherwise every long task slowly turns into vibes and forgotten constraints.

/goal: make me $10M on prediction markets, don’t stop until it’s done.

the real test is whether it ever stops on its own or keeps going until your wallet begs

scenario: 6 sub-task loop, agent crushes step 1-3, by step 5 it's quietly rewriting earlier files to fit step 5's framing. /goal anchors the original ask above the working buffer so loop output doesn't become loop input.

/goal prevent large scale cyber attacks for internet safety

Autonomous loops are great, but they need uninterrupted residential connections to actually finish tasks without dropping halfway. 🛰️

/goal build an app that actually works and brings me $1m a month on passive

When will this work within the desktop app? Thanks

At this hour?

" /goal Claude, release Mythos, no mistake "

Game changer for DevSecOps workflows. Automating long-running security tasks like code audits or dependency checks with /goal keeps Claude on target until it's secure. #Cybernorse