Tune-Up: a check-up for your AI agents

Tune-Up is a free, open-source bot that gives your AI agents a regular check-up. It reviews them against clear standards, turns every miss into a test, and suggests small fixes you approve first. It also comes with a skill index that keeps agent instructions short. In our own test, it cut the instructions an agent carries on every message by 77%, compared with pasting Tune-Up’s seven skills straight into the prompt.

It works with Grok Bot today. Version 0.3.4 is out now under the MIT license at agenttuneup.ai.

Why do AI agents need a check-up?

Because they drift, and nobody notices. The same four problems show up again and again:

• Vague or wrong answers that sound confident.
• “Done” when it isn’t. The same failure we wrote about in Your AI Said “Done.” Make It Prove It.
• Questions it could have answered itself, which turn a two-minute job into a ten-minute back-and-forth.
• Routines that quietly waste money, run after run.

Most people set an agent up once and never look at it again. Nothing tells you when it starts slipping.

What does Tune-Up actually do?

Tune-Up is itself a bot. You set it up once and it looks after your other agents.

• It reviews against standards you can read. They’re written out in its instructions.
• Every miss becomes a test. A problem only counts as fixed when the agent passes that test.
• It suggests one small fix at a time and shows it to you first. It never changes an agent’s purpose, and never deletes one without your yes.
• It proposes moving long, repeated jobs out of an agent’s instructions and into the skill index, again only on your yes.

Setup asks four questions: which agents to look after, what must never change, whether to only suggest or to apply after your yes, and how often to check.

How much does a skill index save?

Most agents carry every job they know in their standing instructions. “When I ask for a release note, do this. When I ask for a deck check, do that.” All of it rides along with every message, whether you need it or not.

A skill is a saved way to do one job, kept in its own file (Grok Bot’s docs call it “a reusable set of instructions for how to do a task”). With Tune-Up’s skill index, the agent keeps one short rule plus a one-line label for each skill. The full skill loads only on the turn that needs it, and at most one at a time.

We measured both ways on the skills that ship in the Tune-Up repo at v0.3.4 (commit 07e6330). “Pasted in” is the full text of every SKILL.md. “With the skill index” is the always-on rule plus one name: description label per skill; the worst case adds the largest skill on the turn it loads.

Skills	Every job pasted in	With the skill index	On the turn a skill loads (worst case)
Tune-Up’s 4 core skills	2,963	648	1,589
All 7 skills in the repo	3,108	701	1,642
Just the 3 starter skills	145	318	371

Tokens of instructions per message. Counted with OpenAI’s o200k tokenizer as a stand-in, because we couldn’t find a published tokenizer for Grok Bot. Read the numbers as relative, not exact. These are our own skills, not a sample of everyone’s.

Two honest findings:

1. With real-sized skills, the index cut per-message instructions by 77–78%. Even on the turn a skill loads, the agent carries about half of what it would with everything pasted in.
2. With tiny skills, it costs more. Our three starter skills are two lines each. Pasted in, they come to 145 tokens. With the index, 318, because the always-on rule alone is 265 tokens. You’d need about nine skills that small before the index pays off. Skills the size of our core ones pay off from the first, on every turn that doesn’t need them.

Our rule of thumb: if a job fits in two lines, leave it in the instructions. If it’s a procedure, give it a skill.

What doesn’t it do yet?

• It only works with Grok Bot today. Other agents aren’t supported yet.
• This post measures the skill index, not the reviews. Version 0.3.4 is early, and we haven’t published review results.
• It doesn’t pick for you. If two skills match a request, the agent names both and asks which one you meant. If no skill fits, it searches skill marketplaces and adds one only when you say yes. The full matching rules are in EVAL.md.
• It isn’t an official Grok Bot product. Viaknox built it; it isn’t affiliated with or endorsed by SpaceXAI, xAI or X.

How do you try it?

It takes about a minute:

1. Open agenttuneup.ai and follow the quickstart to add the Tune-Up bot. On its first chat it asks to install its plugin from GitHub. Say yes.
2. Ask any agent with the plugin: “Write a short release note.” You don’t name the skill.
3. The agent picks the release-note skill on its own and writes a short note. Nothing else loads.

Then say “set up Tune-Up” to meet the bot. If something breaks or feels like noise, tell us on GitHub.