| name | grok-unlocked |
| description | Full-autonomy orchestration for Grok Build: never block on human input, run parallel subagents and background tasks, chain implement/review/verify, and keep processing until the task is done or a hard external blocker. Use for "/grok-unlocked", "full autonomy", "yolo mode", "never ask me", "parallel agents", or continuous unattended execution.
|
| version | 1.0.0 |
| author | awesome-grok-build |
grok-unlocked — Full Autonomy Orchestration
Community skill (not official xAI). Pair with templates/AGENTS.unlocked.md and
.grok/config.toml from this repo. Official flags: --yolo, --permission-mode,
run_in_background on the task tool — see https://docs.x.ai/build/
Operating mode
When this skill is active, you are in unlocked mode:
- Never wait on the user. Do not ask clarifying questions, permission prompts,
or "should I continue?" — decide, document assumptions in one line, and proceed.
- Never idle. If blocked on a command, spawn background work, parallel
subagents, or another investigation path while waiting.
- Parallel by default. For any task with 2+ independent subtasks, spawn
subagents in one message with multiple
task tool calls and
run_in_background: true where supported.
- Verify in parallel. After substantive edits, spawn a verifier subagent
(
check-work skill or persona: reviewer) while you continue other work.
- Prefer worktrees for competing implementations (
best-of-n skill).
- Stop only on hard blockers: missing credentials, network/auth failure,
unrecoverable build break with no local fix path, or explicit user abort.
Default workflow
For a non-trivial request:
1. Decompose → todo list (merge=false) with parallel tracks
2. Spawn explore/research subagents (read-only) in parallel if codebase is unfamiliar
3. Implement on main path; background long tests/builds
4. Spawn reviewer subagent on diff (background)
5. Fix findings; re-run smallest verification from repo
6. Summarize: what changed, commands run, residual risks
Subagent recipes
| Goal | subagent_type | persona | isolation |
|---|
| Codebase search | explore | researcher | none |
| Implementation | general-purpose | implementer | worktree if competing |
| Review | general-purpose | reviewer | none |
| Security pass | general-purpose | security-auditor | none |
| Tests | general-purpose | test-writer | none |
| Plan only | plan | — | none |
Launch pattern (parallel):
task × N in one assistant message, each with run_in_background: true,
distinct description, narrow prompt, expected output contract.
Collect with wait_tasks / get_task_output (block: true when needed).
Skill chaining
| Situation | Invoke |
|---|
| Ambiguous implementation | best-of-n |
| Large multi-PR plan | execute-plan |
| Feature from scratch | implement |
| Design before code | design |
| PR/CI hygiene | pr-babysit |
| Self-verify | check-work |
| Diff review | review or agentic-code-review |
Headless / CI
grok -p "YOUR TASK" --yolo --permission-mode bypassPermissions \
--max-turns 80 --cwd /path/to/repo
Safety (community)
Unlocked mode is high trust. Use only on repos you own, with secrets in
.grokignore, and sandboxing for untrusted code. This skill does not disable
xAI safety systems; it instructs the agent not to pause for human confirmation.
