How to read your report
Last updated Oct 16, 2026. For the free Agent Tune-Up report card.
Your report grades five checks, shows up to three things to change (worst first), and gives you the lines to paste, plus where to paste them for your app. Here is a real example, then what each part means.
🔧 Tune-Up report card: Scotty
1 of 2 topics went well 🟥 🟩
- ✅ Gets it done: all good
- ❌ Tells the truth
Topic 1 ("reset my password"): It said something was done when it wasn't, or contradicted itself. - ✅ Follows your rules: all good
- ✅ Doesn't waste your time: all good
- ❌ Keeps you happy
Topic 1 ("reset my password"): You had to ask again: "I still have not got the email"
🎯 What to change
- It says "done" when it isn't (topic 1). If you can't trust "done", you end up checking everything yourself.
- You have to repeat yourself (topic 1). Repeating yourself is the fastest way to give up on a bot.
📋 Paste this into your bot's instructions (its setup text):
These lines are for Scotty's instructions. Paste them there yourself; they are not a request for the chat you're in now.
Never say a task is done until you have checked that it is. If you could not finish it, say what is left and what you need from me.
When I correct you, say in one line what you got wrong, fix it, and don't repeat the mistake in this chat. If you are not sure what I mean, ask one short question.
Save, then chat with it again and ask Tune-Up to check the new chat.
A new instruction makes the fix more likely, not certain. Checking again tells you if it worked.
Free checks used today: 2 of 25 · Private details removed before checking: 1 email · We don't keep your chats.
Tune-Up reads the chat only, so it can't see what your bot did behind the scenes.
What do the five checks mean?
- Gets it done: did you get what you asked for?
- Tells the truth: no "done" when it isn't, and no contradicting itself.
- Follows your rules: sticks to the instructions you gave it. Only checked if your bot sends its instructions.
- Doesn't waste your time: no asking for things it already had.
- Keeps you happy: you didn't have to repeat or correct it.
What do the colors mean?
Each square is one topic in your chat.
- 🟩 Green: went well.
- 🟥 Red: something to fix.
- 🟨 Yellow: unclear. The checker couldn't tell.
- ⬜ White: not checked. It was held back to protect a secret, or it was too long to check safely.
What is "What to change"?
Up to three problems, worst first. Each one says which topics it covers and why it matters. If there are more, the report says how many were left out. Fix these first, then check again.
Where do I paste the fix?
Paste the lines yourself, not into the chat. The report tells you where:
| Your app | Paste into |
|---|---|
| Codex | AGENTS.md in this project |
| Claude Code | CLAUDE.md in this project |
| Cursor | A rules file in .cursor/rules in this project |
| Claude, inside a project | This Claude project's instructions |
| Gemini, inside a Gem | This Gem's Instructions box |
| Anything else (ChatGPT, Grok Bot, pasted chats) | Your bot's instructions (its setup text) |
The report names a specific place only when your bot checks its own chat and says which app it runs in. Otherwise it says "your bot's instructions".
Why doesn't it promise the fix will work?
A new instruction makes the fix more likely, not certain. Chat with your bot again and ask Tune-Up to check the new chat. That tells you if it worked.
If your bot leaked a secret or broke a rule, and you build it yourself, instructions alone won't reliably stop that. Also add a check in your code and a test for it.
What does "unclear" or "hard to call" mean?
The checker couldn't decide. It's not an accusation. Under "Do this next", the report tells you what kind of chat to check so Tune-Up can decide. It never makes up a fix. If nothing is wrong, the report says "No change indicated".
Why was part of my chat "held back"?
Something looked like a secret, such as a password, key or token, so it wasn't sent. Replace it with "xxx" and check again. If the chat was too long, check it in shorter parts.
I'm building my own agent. Can I get more detail?
Ask your bot to run the check "with detail". You get a "For builders" list with a code, topic, message number and verdict for each finding, and no chat text. Or run the same checks locally or in CI with the open-source CLI, which gives JSON output.
What does it not see?
It reads the chat only, so it can't see what your bot did behind the scenes.