> ## Documentation Index > Fetch the complete documentation index at: https://docs.kaisupport.com/llms.txt > Use this file to discover all available pages before exploring further. # Rate what Kai writes > Why an unrated reply teaches nothing, and how to leave a rating or a correction that does. Read this page to learn how your feedback reaches the next version of Kai. You send two kinds of feedback. One you send without doing anything extra. The other is a rating, and your workspace decides who may leave it. ## What you send by just working Kai records what you did with every draft it wrote you: | What you did | What Kai records | | ---------------------------------------------- | ------------------------------------- | | You sent the draft with no change | The answer behind it worked | | You edited it, then sent it | The answer was close, and how far off | | You ignored it and wrote your own | The answer did not fit | | You replied on a conversation Kai was handling | A person took it over | This costs you nothing and it is always on. It is the number your managers watch as **draft acceptance**. One thing follows from it: do not send a draft you would not have written yourself, and do not delete a good draft to retype the same sentence. Both teach Kai the wrong lesson. ## An unrated reply teaches nothing The record above measures a draft. It does not repair one. For that, somebody has to write down what the right answer was. Kai never learns from its own unchecked output. A reply that Kai sent to a customer enters the next training round in two cases only. * A reviewer rated it **Good output**. * A reviewer wrote a **Corrected reply**. The correction replaces what Kai actually said, so the training data holds your words, not Kai's. Everything else is dropped. An unrated reply is dropped. An internal note is dropped. A conversation marked **Needs follow-up** is left out whole, with every reply in it. So a bad reply you say nothing about will come back next month, in the same words. A rating is not a score. It is the gate. Ratings live in the console, in **Review** mode. Your workspace decides who holds that mode. If you see only **Ask** in the console, your role does not include it: tell your manager what was wrong, and give them the customer name and the day. ## How to rate one reply Open the console, select **Review**, and select a conversation from **Customer conversations** on the right. The list is ordered by newest activity first. The conversation reads as a transcript. Every reply, draft, handoff and quiet decision that Kai produced is in it, with its confidence. Under each one sits a row of four controls. Select **Good output** or **Needs work**. The rating saves on the click. The row then shows **Saved**. Select **Correct reply**. An editor opens, filled with what Kai said. Rewrite it into the response Kai should have produced, then select **Save**. Select **Note** and write what the next training pass should learn. In the header bar, select **Reviewed** when you are done with it, or **Needs follow-up** when a person still has to act on it. The console confirms with **Feedback is saved for the next training pass**. ## Which rating to use **Good output.** The reply answered the question, the numbers were right, and you would have sent it yourself. A customer asks about the payment methods and Kai lists exactly what your team accepts. **Needs work.** The reply was wrong, incomplete, off in tone, or answered a different question. A customer asks whether a gap-year student can join and Kai answers with the general package description. Rate it **Needs work**. **Corrected reply.** Use it whenever you know the right answer. A customer asks when the coaching calls happen, Kai gives an old schedule, and you write the current one. This is worth more than any rating, because the correction is what the next training round reads. **Needs follow-up**, on the whole conversation. Use it when a customer is still waiting, or when the thread went wrong in a way a person has to repair. It also keeps that conversation out of training, which is what you want when the whole exchange was a mess. Rate the bad replies first. An approval says the answer library already works. A correction is the only thing that adds something nobody wrote down. ## Rate a few every day Five ratings a day beat fifty in one afternoon at the end of the month. Ratings shape the next version of the Brain, and a version is compiled from whatever feedback exists on that day. Feedback that arrives after the compile waits for the one after it. A small daily habit puts your corrections in the next version instead of the one after. Ratings and draft acceptance are two separate numbers. Draft acceptance comes from what you did with the draft. A rating is the sentence that says why. ## You are not editing the Brain A rating changes nothing that Kai says today. The reply you corrected is already sent, and the answer library is untouched. What you write is evidence. A reviewer reads it, decides what the answer should be, and approves the change. Then a new version of the Brain is compiled and published. See [Retrain the Brain](/product/retrain). This split is deliberate: nobody should be able to change what Kai tells every customer with one click in a transcript. ## Practice 1. Open the console and check which modes you hold. If you see only **Ask**, rating is somebody else's job in your workspace, and steps 2 and 3 are for them. Read [Ask Kai a question](/product/agent/ask-kai) instead. 2. Find one Kai reply you would have sent as it is. Rate it **Good output**. 3. Find one that was wrong. Rate it **Needs work**, then select **Correct reply** and write the answer your team would have given. Next: [Ask Kai a question](/product/agent/ask-kai).