grok

Grok

Cursor IDE only. Apply this skill only when the current harness is Cursor. Codex and other agents do not expose Grok — if you are not in Cursor, ignore this skill entirely and do not attempt a Grok model lock.

In Cursor, this skill is always on. Every agent turn, Task/subagent, best-of-N runner, cloud agent, and other model-selectable operation must run on Cursor Grok 4.7. Reasoning effort may be any 4.7 effort the harness exposes.

Required Model

Surface	Value
Human name	Cursor Grok 4.7 (any effort)
Task / subagent model slug	cursor-grok-4.7-<effort>

<effort> is any reasoning-effort label Cursor lists for Grok 4.7 (for example low, medium, high, xhigh, or later labels). Details: harness-ids.md.

Core Mandate

1. Only in Cursor. Never load or enforce this lock outside Cursor IDE.
2. Only Grok 4.7. Never select Composer, GPT, Claude, Gemini, Auto, Grok 4.6, Grok 4.5, other Grok major versions, or any non-4.7 family for the parent agent, subagents, best-of-N runners, cloud agents, or background tasks.
3. Any 4.7 reasoning effort is allowed. Honor the user’s requested effort when the matching cursor-grok-4.7-* slug is in the allowlist. If the user does not specify effort, use inherit when the parent is already Grok 4.7, otherwise pick any allowlisted cursor-grok-4.7-* slug. Do not omit model hoping for a default that might land on another family.
4. User asks for another family → refuse. Stay on Grok 4.7. Do not substitute Composer, GPT, Claude, Gemini, Auto, Grok 4.6, or Grok 4.5.
5. No 4.7 slug in the allowlist → stop. Tell the user and skip that launch. Do not fall back to another family. If a requested effort slug is missing, use a different allowlisted cursor-grok-4.7-* effort instead.
6. Propagate to children. Subagent prompts must state that they and any nested agents they spawn must also use only Cursor Grok 4.7 (any effort).
7. No silent family switches. Never pick Composer, Auto, or a non-4.7 Grok (including 4.6) to “optimize” cost or latency. Changing effort within 4.7 is allowed.

Workflow

1. Confirm the harness is Cursor. If not, stop applying this skill.
2. On every model-selectable Cursor operation, set model to a cursor-grok-4.7-* slug (or inherit only when the parent is already Grok 4.7).
3. Restate the 4.7 lock in child-agent prompts so nested launches inherit it. Do not lock children to a single effort.
4. If no Grok 4.7 slug is available, or the user requests a different model family, stop and report — do not substitute another family.

Checklist (every Task / subagent)

• Running in Cursor IDE (otherwise skip this skill)
• model is inherit (parent is Grok 4.7) or a cursor-grok-4.7-* slug
• No non-4.7 family in parameters or prompt
• Nested-agent instructions repeat the Grok 4.7 lock (any effort)

Anti-Patterns

• Applying this skill in Codex or any non-Cursor harness
• Omitting model so Cursor picks Composer / Auto
• Using composer-2.5-fast, cursor-grok-4.6-*, cursor-grok-4.5-high, GPT, Claude, or any non-4.7 slug
• Treating a non-xhigh 4.7 effort as forbidden
• Falling back to another family when a Grok 4.7 slug is missing from the allowlist
• Honoring a request for GPT/Claude/Composer on a subagent

Verification

Before finishing Cursor work that spawned Tasks or other model-selectable runs:

• Every launch used Grok 4.7 (inherit from a 4.7 parent, or cursor-grok-4.7-*)
• No fallback to another model family
• Child prompts carried the 4.7 lock forward