Claude Code supports /goal QA and marketing runs with proof-based success checks

• Anthropic positioned /goal as a completion condition for long Claude Code runs, where Claude keeps working until a stated condition is met, according to ClaudeDevs' launch thread and ClaudeDevs' example prompt.
• The core prompting pattern is tighter than a normal todo: minchoi's framing boiled it down to a clear goal, something measurable, proof, and limits.
• Early examples already split beyond coding, with minchoi's QA demo showing a lint-and-test pass and shannholmberg's marketing thread mapping the same pattern onto content audits, SEO rewrites, swipe files, and outreach research.
• Anthropic's own thread also tied /goal to adjacent long-run controls, with ClaudeDevs on the Ralph loop describing the stop-check loop, while ClaudeDevs on /loop, ClaudeDevs on /schedule, and ClaudeDevs on auto mode filled in the rest of the harness.

You can read the official Claude Code /goal docs, watch ClaudeDevs' demo thread run a test-and-lint target to completion, and skim shannholmberg's marketing missions for a creator-side version of the same idea. There is also a small product-surface wrinkle, because LLMJunky's Codex App walkthrough showed /goal hidden behind a Cmd/Ctrl+J command palette path rather than an obvious inline command.

Anthropic described /goal in plain terms: set a completion condition, and Claude keeps going until it can satisfy it.

The most concrete demo was a QA chore. In minchoi's video post, the prompt is simply /goal all tests pass and lint is clean, which is exactly the kind of pass-fail target an agent can check without asking for interpretation.

Anthropic added one mechanical detail in ClaudeDevs on the Ralph loop: every time Claude tries to stop, it checks the condition against the transcript. If the condition is not met, it keeps going. If it is, Claude returns a "Goal achieved" summary.

The best unofficial summary came from minchoi's checklist post, which reduced a good /goal to four parts:

• Set a clear goal
• Make it measurable
• Show proof
• Add limits

That matches Anthropic's surrounding controls. ClaudeDevs on stop hooks said a stop hook can gate completion on external checks like a test suite or CI endpoint, while ClaudeDevs on auto mode said long-running tasks work best when Claude does not have to stop and wait for a human.

Anthropic also grouped /goal with two adjacent commands in the same thread:

• /loop for repeated iterative runs, per ClaudeDevs on /loop
• /schedule for cadence-based runs like nightly tests or weekly cleanup, per ClaudeDevs on /schedule

The interesting twist is that the pattern already escaped coding. In shannholmberg's thread, /goal becomes a reusable structure for creator and marketing work, with the same insistence on checkable outputs.

The five example missions were:

1. Content audit plus restock
2. Competitor swipe file
3. SEO plus content gap audit
4. ICP research at scale
5. Voice plus style consolidation

The useful part is not the category list, it is the success condition attached to each one. shannholmberg's image specifies outputs like 20 queued drafts, a 100-example swipe file, scheduled republishes, a 200-row outreach sheet, or a reusable voice.md, which turns fuzzy "help me with marketing" asks into deliverables the agent can verify.

One smaller wrinkle came from LLMJunky's Codex App steps, which claimed /goal was available in the Codex App but not directly discoverable from the main thread view.

The steps in that post were:

1. Press Ctrl/Cmd+J in any thread
2. Type codex and press Enter
3. Type /goal [set your goal]

That is a separate product surface from Claude Code, but it adds one new practical detail to the story: goal-based runs may exist before they are obvious in the default UI.