Grok Bot Skills + Which AI Models To Draft, Test And Port Them With, Using Real GoldieBench Scores

Grok Bot skills decide how a task gets done, but they don't let you choose the model doing it, so the model question moves to where you draft and test them.

The Grok Bot FAQ is clear: there is no model picker, and Grok Bot picks the backend model for each task.

So I read the live GoldieBench pages on 3 November 2026 to see which models are worth using around your skills.

Want your own routing mapped? Book a free SEO strategy session with my team.

Grok Bot picks the model, so skills are your real lever

Inside Grok Bot, you control the instructions, not the engine.

A skill holds the steps, decision rules, expected output and approval limits, per the skills docs.

That makes the skill text itself the thing worth getting right.

The benchmark question becomes: which model helps you write, check and port that text best?

What GoldieBench measures

The GoldieBench leaderboard showed 1310 live demos, 50 tasks and 33 models when I read it.

Every model gets the same prompt, one shot, and the result is scored 0 to 10.

It tests whether a model can hold a long spec and ship what it promises.

Medals show per-task rank, so a model with many golds wins head-to-head briefs often, not just on average.

It doesn't test skill files, tool choice or approval handling directly, so read the scores as a proxy for following detailed instructions.

The top seven on the board, read 3 November

All seven rows come from the leaderboard home page.

The spread from #1 to #7 is exactly half a point.

Grok: 8.09 average, 43 of 47 tasks scored

The Grok model page lists 47 tasks tested, an 8.09 average and medals of 5 gold, 1 silver and 1 bronze.

Its best single build is Twilightvale at 9.5, and its golds include Landing Page and Webos at 9.0 each.

The page lists its best use as workflows that need live X context, with a 256,000-token context window.

That's the family Grok Bot sits closest to, which makes it a sensible first test for skills that touch X content.

Its lower scores sit in visuals, with Aurora and Matrix at 7.0, so visual output steps are better drafted elsewhere.

Grok 4.7 on twenty skill-infused builds

The Grok 4.7 model page says it was benched on twenty skill-infused game builds, each played through before the judge scored it.

It averaged 7.15 on those briefs, where Grok 4.6 averaged 5.95, and won 14 of the 20 head-to-heads.

Its best one-shots were neoncity and racing at 8.7 each, and arcade at 8.6.

Two of the 20 builds died on load, with crypt at 2.3 and voxelcraft at 2.2 on the deep dive.

One detail to note: the Grok 4.7 deep dive, refreshed 2026-10-27, gives board rank #18 of 26, while the live leaderboard listed it at #17 on 3 November.

Which model drafts which part of a skill

The docs list six parts a useful skill should state. Here's how I'd split the drafting by benchmark strength.

Paste the drafted text into the Bot, ask it to save the skill, then test it on a safe example.

Drafting outside the Bot costs one extra step, but it lets you compare two models on the same brief.

Give both the same finished task, ask each for the six-part skill, and keep the clearer one.

That mirrors the bench itself: same prompt, one shot, judged on the result.

Plugin and template skills don't change the model

Since v0.71.2 a Bot can use up to 200 skills from installed plugins, and since v0.75.8 a newly installed plugin's skills work right away, per the Grok Bot changelog.

Since v0.77.0, template skills show under Private skills labelled From template.

None of these expose a model choice; they widen what the Bot can do, not which engine runs it.

Plugin setup lives on our Grok Bot plugins page.

Port your skills to Hermes to pick the model yourself

If a skill needs a specific model, move it somewhere you choose one.

The Hermes import guide says hermes import-agent grok-bot prints a message for each Bot to export itself, and imported skills land in skills/grok-bot-imports/.

In Hermes you then point that skill at whichever leaderboard model fits the job.

Keep the Grok Bot copy for the work that needs its connected apps.

What the scores don't tell you about skills

A one-shot build score says a model can follow a long brief.

It doesn't say whether a Bot will stop at an approval point, pick the right one of 200 plugin skills, or handle a missing data source.

For that, the docs' own method matters more: test on a safe example, review what it did, and keep sending, buying, deleting and publishing behind approval.

On Enterprise, audit logs can also record which skills a Bot read, which is the real check on skill use.

Think of it the way the bench works: same brief, same rubric, every run.

Give your skill the same safe input twice and compare the two outputs before you trust it with live data.

If the runs differ, the skill text is too loose, and no model swap inside Grok Bot will fix that.

A five-step routing checklist

1. Does the skill run inside connected apps? Keep it in Grok Bot.
2. Draft long step lists with a top-three model, then save the skill in the Bot.
3. Need a fixed model? Export the skill to Hermes.
4. Write every approval rule yourself.
5. Check your plan: Premium, Super and Ultra include Grok Bot, and Plus doesn't, per the plan docs.

Want the Agent OS that runs these models side by side? Join the AI Profit Boardroom.

Or grab a free strategy call and route your first skill with us.

Also On Our Network

• Agent Operatorsthe five-step operator setup for saving, enabling and publishing Bot skills
• AI Income Deskwhich repeat jobs a small business should turn into Bot skills first
• AI Tool Verdictan eight-feature scorecard rating Bot skills out of ten
• agentos.guidefive copy-paste prompts for organising and backing up Bot skills
• aiprofitboardroom.comthe complete Grok Bot skills guide on the AI Profit Boardroom blog

Draft with a top-scoring model, save and test in the Bot, and port to Hermes when a skill needs a fixed engine. For help choosing, book a free strategy session and match a model to each of your Grok Bot skills.