Verifiable run-until-done objectives for Grok Build /goal. Patterns, a verifier, and the goal-audit CLI.
Goals are for bounded, verifiable finish work. Use this decision guide before /goal.
flowchart TD
A[New work item] --> B{Exploratory / spike?}
B -->|yes| P[One-shot prompt or chat]
B -->|no| C{Recurring discovery?}
C -->|yes| L["/loop + STATE.md"]
C -->|no| D{Human approves every step?}
D -->|yes| G1[G1 goal with human gates]
D -->|no| E{Verifiable done condition?}
E -->|no| S[goal-scoper first]
E -->|yes| G["/goal run-until-done"]
You do not know what “done” looks like yet. Use a normal prompt; capture learnings in a doc, not GOAL.md.
“Understand the codebase” has no objective gate. Use chat or a time-boxed loop, not a goal.
Use G1 assisted goals with explicit human checkpoints — or skip autonomy and pair program.
Security scans, inbox triage, dependency bumps on cadence → loop-engineering.
Subjective goals (“make it nicer”) fail verifiers. Run goal-scoper to add npm test, coverage %, or grep gates.
Pause one goal or serialize. See multi-goal.md.
| Work type | Recommendation |
|---|---|
| Spike / prototype | Prompt |
| Daily triage | Loop |
| Finish one ticket | Goal |
| 5 related tickets | 5 sequential goals |
| “Fix everything” | goal-scoper → split |
Next: pattern-picker.md · golden path