# When you edit, delete or replace a draft

A draft the assistant wrote sits in your mailbox, in the Drafts folder. What
you do with it is entirely up to you — and draften learns from it. Here is
what exactly it watches and what follows.

## What draften watches

At every mail check draften looks at the drafts that have no outcome yet:
whether the draft is still in Drafts, whether you sent it, and if so, how
much of it stayed. It needs no language model for that; it just compares the
texts.

| What you do | Status in Drafts | What follows |
|---|---|---|
| send the draft as it is | sent unchanged | the draft was right; nothing more is evaluated |
| edit it and send it | edited (at least half kept) | draften compares both versions and evaluates why you changed it |
| write the reply your own way | replaced by your own reply | less than half of the draft stayed; it is evaluated too |
| delete the draft | deleted | it is evaluated whether a reply was due at all |
| leave it lying | expired | after 14 days unsent draften stops watching it |

A deleted draft counts as deleted only after a while, not at once: when you
delete it and write the reply yourself, that takes time. How long draften
waits is in Settings → Behaviour and limits (2 hours by default).

## Content evaluation

For an edited, replaced or deleted draft draften compares what the assistant
wrote with what went out and gives a **content evaluation** with a short
reason. It is in the draft's detail and in a column of the Drafts list.

| Evaluation | What it means |
|---|---|
| correct | the content was right; only a detail differed |
| form only | the facts were right; you changed the way of writing (tone, length, greeting) |
| wrong fact | the draft claimed something untrue |
| missing information | the draft left out something important |
| wrong diagnosis | the assistant understood the problem differently than it was |
| other question | it answered something else than the customer asked |
| should have escalated | it should have handed the matter to a colleague |
| should not have replied | the mail should not have been answered at all |
| other circumstances | you knew something the assistant could not (a phone call, an agreement) |
| unclear | the change does not show why |

**Other circumstances** do not count towards the success rate on the Overview
and nothing is learnt from them — it was not the assistant's fault.

## What the assistant learns

The evaluations turn into **proposals** in the Learning section:

- repeated edits of form give a **lesson** — a general rule of writing, such
  as "When moving a meeting, offer two new dates";
- a corrected fact gives a **knowledge correction** — what the draft claimed
  and what is right.

A proposal comes up for approval once it has been seen at least twice. The
assistant uses none of it until you approve it. See the guide
[Learning: lessons and fact corrections](/en/learning).

If you do not want the assistant to learn from your edits, switch off
**Learn from my edits of the drafts** in Settings → Behaviour and limits.
Drafts are still watched, only not evaluated, and nothing new is proposed.

## Why it matters

The more drafts you send unchanged or with a small edit, the better the
assistant writes. You see the success rate on the [Overview](/en/overview),
in the Drafts card. When the same mistake keeps coming, do not wait for a
lesson — change the [instructions](/en/instructions) or the
[knowledge sources](/en/knowledge-sources) directly.