goal-command

name	goal-command
description	User wants Claude to keep working on a task without stopping to check in; user wants to set a completion condition and walk away; user asks about a...

/goal — Autonomous Task Completion

When to activate

User wants Claude to keep working on a task without stopping to check in; user wants to set a completion condition and walk away; user asks about autonomous operation or keeping Claude running until a specific outcome is reached.

When NOT to use

Simple single-step tasks where one response is sufficient; tasks where the user wants Claude to pause and confirm after every action; interactive debugging sessions where back-and-forth is the point.

Instructions

Syntax:

/goal <completion condition>

The condition is evaluated after every assistant turn. Claude keeps working — writing code, running commands, seeing failures, adjusting — until the condition holds, then stops and reports.

Writing good conditions:

Natural language works. The condition should be observable and unambiguous:

• All tests pass — Claude runs the test suite, fixes failures, reruns, until green
• The PR is created — Claude finishes the work and opens a PR
• The migration runs without errors — Claude applies the migration, checks for errors, fixes schema issues
• tsc --noEmit exits 0 — Claude resolves TypeScript errors until the compiler is clean
• CHANGELOG.md exists and has today's date — Claude writes the changelog file

Bad conditions to avoid:

• Subjective: "looks good", "is clean" — not verifiable by Claude
• Open-ended: "keep improving the code" — no stopping condition
• Time-based: "run for one hour" — not an outcome

Combine with effort level for maximum autonomy:

/goal All tests pass
/effort xhigh

Interrupting: Send any message to interrupt, or delete .claude/goal to cancel. The goal state persists across context compactions — Claude remembers the goal even after the context window compresses.

Background sessions: Works with claude --bg. Set the goal, close your terminal, come back when it's done.

What happens in each turn:

1. Claude takes actions (edits files, runs commands)
2. Evaluates: does the condition hold?
3. If no — continues
4. If yes — stops and reports what was done

Example

/goal All TypeScript errors are resolved and tsc --noEmit exits 0

Claude runs tsc --noEmit, reads the error list, fixes each error, runs again, sees remaining errors, fixes those, runs again — loop continues until zero errors. Then stops and reports: "Resolved 14 TypeScript errors across 6 files. tsc --noEmit exits clean."

---