grok-build – Agents & Automation skill for AI agents
Delegate well-specified implementation tasks to xAI's Grok Build CLI running headlessly while the orchestrating agent plans, writes task specs, reviews every diff, and owns the result.
Install the grok-build skill with one command
Pick the tool you use, copy the command and paste it in your terminal. The skill lands in that tool's skills folder and is ready in your next session.
Claude Code
mkdir -p ~/.claude/skills && curl -fsSL https://skill.dn1nguoi.com/z/grok-build-e7fd7.tgz | tar xz -C ~/.claude/skills
Call it in Claude Code: /grok-build – then describe the task you want done.
Codex
mkdir -p ~/.codex/skills && curl -fsSL https://skill.dn1nguoi.com/z/grok-build-e7fd7.tgz | tar xz -C ~/.codex/skills
Call it in Codex: $grok-build – then describe the task you want done.
Hermes Agent
mkdir -p ~/.hermes/skills && curl -fsSL https://skill.dn1nguoi.com/z/grok-build-e7fd7.tgz | tar xz -C ~/.hermes/skills
Call it in Hermes Agent: /grok-build – then describe the task you want done.
DeepSeek
mkdir -p ~/.deepcode/skills && curl -fsSL https://skill.dn1nguoi.com/z/grok-build-e7fd7.tgz | tar xz -C ~/.deepcode/skills
Call it in DeepSeek: /grok-build – then describe the task you want done.
SKILL.md content
Grok Build Orchestration
When to Use
- Use when delegating a well-specified implementation task to xAI's Grok Build CLI running headlessly
- Use when executing a Markdown implementation plan task-by-task with a diff review after each task
- Use when the user says "use grok", "grok build", "have grok implement", or "send to grok"
The coding assistant is the orchestrator: it plans, writes self-contained task specs,
dispatches them to Grok Build headlessly, reviews every diff, and owns the final result.
Grok is the fast, cheap executor. Full CLI details and verified behaviors: references/cli.md.
Safety Gate
Before every dispatch, show the user the exact task specification that will be sent to xAI,
the target worktree, and the permission mode. Obtain explicit approval to disclose that text
and to let Grok edit the scoped worktree. Never include secrets, proprietary source, customer
data, or credentials in a task specification. Do not run grok update, --always-approve,
or a destructive recovery command without separate, explicit approval.
When to delegate vs keep with the orchestrator
| Delegate to Grok | Keep with the orchestrator |
|---|---|
| Plan tasks with clear acceptance criteria | Ambiguous requirements, architecture decisions |
| Boilerplate, scaffolding, CRUD | Deep cross-file debugging |
| Mechanical refactors | Security-sensitive code |
| Test writing from clear specs | Anything touching production infrastructure |
| UI components from mockups/specs | Tasks where writing the spec ≈ doing the work |
When in doubt, keep it with the orchestrator.
A prompt/file list is not a security boundary. --always-approve also permits shell commands: verify OS/container confinement, minimal environment credentials, network policy and budget before dispatch; if unavailable, do not run this headless path. Review all changed/untracked files and observed side effects, not only spec-listed paths.
Session preflight (once, before the first dispatch)
1. grok update --check --json — if updateAvailable is true, tell the user. Run
grok update only after explicit approval, then confirm with grok --version.
2. grok models — if it errors or reports logged out, STOP and ask the user to run
grok login.
Per-task loop (sequential — the default)
1. Spec. Write a self-contained task file (template below) to a temp directory
OUTSIDE the target repo — the harness scratchpad if one is available, else the OS
temp dir. Never write it inside the target repo. Grok has zero conversation context:
no one-liner prompts, ever.
- POSIX:
mkdir -p "${TMPDIR:-/tmp}/grok-specs", then writetask.mdthere. - Windows (PowerShell):
New-Item -ItemType Directory -Force "$env:TEMP\grok-specs",
then write task.md there.
2. Clean state. No uncommitted *source* changes — commit or stash first, so the
post-run diff is exactly Grok's work. Ignore build artifacts (__pycache__, dist/,
etc.); if they show in git status, they're usually just un-gitignored, not your
concern. Never dispatch on a dirty source tree.
3. Dispatch.
POSIX:
grok --prompt-file <task-file> \
--output-format json \
--always-approve \
--max-turns 30 \
--cwd <repo>
Windows (PowerShell) — backtick line-continuation:
grok --prompt-file <task-file> `
--output-format json `
--always-approve `
--max-turns 30 `
--cwd <repo>
Parse the JSON output and save sessionId. (--always-approve is required for
headless runs — --permission-mode acceptEdits silently cancels edits with no
interactive approver. Use it only after the user explicitly approves Grok editing this
exact scoped worktree. See references/cli.md.) For a high-stakes task, add --check
so Grok self-verifies before you review; skip it otherwise (it ~doubles latency).
4. Review gate — non-negotiable.
- Read the diff yourself (
git diff -- <files from the spec>to skip artifact noise):
does it do the task, only the task, and match repo conventions?
- Run the acceptance commands from the spec.
- Pass → commit with a clear message following the repo's convention → next task.
- Fail → ask the user before a fix-up or any reset. Never run
git checkout -- .or
git clean -fd automatically; preserve the diff for review and use a non-destructive
recovery plan unless the user explicitly authorizes otherwise.
Task spec template
# Task: <one-line title>
## Context
- Repo: <path> — <one line on what the project is>
- Conventions: <test runner, formatter, a good example file to imitate>
## Files
- Modify: <path>
- Create: <path>
## Task
<precise description of the change>
## Constraints
- Do not modify any files other than those listed above.
- <other constraints>
## Acceptance criteria
- `<exact command>` <expected result>Executing a Markdown implementation plan
- One plan task per dispatch, in order.
- Check off the plan's task checkboxes (
- [ ]→- [x]) as each task lands and passes
the review gate.
- If the plan explicitly marks tasks as independent, see Parallel dispatch below;
otherwise stay sequential.
Parallel dispatch (opt-in exception, not the default)
Only when a plan explicitly marks tasks independent: dispatch each with
--worktree=<task-slug>, run concurrently, then review and merge one worktree at a
time through the same review gate. Merge conflicts usually eat the savings — prefer
sequential.
Failure handling
| Failure | Action |
|---|---|
| stopReason: "Cancelled", empty text, no diff | Missing --always-approve — retry with it |
| CLI error / timeout | Retry once; then do the task yourself and note the fallback |
| Auth expired | Stop; ask the user to run grok login |
| 2 fix-up rounds exhausted | Preserve the diff, ask the user for a recovery decision, then finish the task manually if authorized |
| Dirty tree at dispatch | Refuse; commit/stash first |
Limitations
- Grok receives the approved task specification; it is a third-party service and should not
receive secrets, proprietary material, personal data, or customer data.
--always-approveallows edits without an interactive approval prompt. It must be limited to
a clean, explicitly approved worktree and never substitutes for the orchestrator's review.
- Model output can be incorrect, insecure, incomplete, or out of scope. Review the diff and
run the acceptance checks before accepting any change.
- This skill does not authorize installations, updates, commits, pushes, deployments, or
destructive cleanup.
Models
Default grok-4.5. Add -m grok-composer-2.5-fast only for trivial mechanical tasks.
Related skills
- helium-mcp – Connect to Helium's MCP server for news research, media bias analysis, balanced…
- helpdesk-automation – Automate HelpDesk tasks via Rube MCP (Composio): list tickets, manage views, use canned…
- hf-mcp – Use Hugging Face Hub via MCP server tools. Search models, datasets, Spaces, papers. Get…
- hf-mem – Hugging Face CLI to estimate the required memory to load Safetensors or GGUF model…
- hierarchical-agent-memory – Scoped CLAUDE.md memory system that reduces context token spend. Creates directory-level…
- skillopt-sleep – Use when the user wants their Claude agent to self-improve from past usage, asks about a…
- hosted-agents – Build background agents in sandboxed environments. Use for hosted coding agents…
- hosted-agents-v2-py – Build hosted agents using Azure AI Projects SDK with HostedAgentDefinition. Use when…