You keep fixing the same mistake because editing a draft usually changes that one email and nothing else. The next draft starts from the same sources and instructions as before, so the same gap produces the same error. A correction only sticks when it lands somewhere the tool reads next time: the document it takes facts from, a standing instruction, a saved answer, or a memory of your edits that you can see and correct.
Why doesn’t editing a draft teach the tool?
Because for most writing tools, your edit is just an edit. You change “30 days” to “14 days” in the draft and send it. The email is right; the tool’s sources are unchanged. Next week, a customer asks the same question, the tool reads the same outdated email or document, or none at all, and writes “30 days” again.
Some tools do learn from edits, by comparing the draft with what you sent. Even then, it matters whether you can see what they learned: a hidden adjustment is hard to trust and impossible to correct.
Where can a correction live?
Four places, each good for a different kind of mistake:
| Where | What it fixes | Example |
|---|---|---|
| The source the tool reads | Wrong or missing facts | Update the refund rule in your help article or facts document |
| A standing instruction | Tone, length, format, phrases to avoid | “Never start with ‘Unfortunately’”; “keep replies under 100 words” |
| A saved answer | A recurring question with a fixed answer | A checked answer to “Can I change my delivery address?” |
| The tool’s memory of edits | Both, if the tool records what it learned and shows it | A readable log entry: “Refund window is 14 days, not 30” |
The most common mistake is fixing a fact in the draft and never in the source. It feels like a correction but changes nothing for next time.
Which kind of mistake goes where?
| Mistake in the draft | Fix it in |
|---|---|
| Wrong price, date, policy or rule | The source: help article, website page, facts document |
| A fact the tool never had | A new source: add the answer to your help center or facts document |
| Too long, too formal, wrong greeting | A standing instruction or the tool’s tone settings |
| Promises something you do not offer | The source (state what you do not offer) and an instruction (“do not promise refunds”) |
| The same good answer rewritten differently every time | A saved answer |
| Answers the wrong question | Nothing to store; read the customer’s email more carefully and regenerate |
How do you make corrections stick in Gmail?
- Facts: keep them in a Drive document and name it in your prompt. Google’s own advice for Help me write: “To ensure factual accuracy, mention a specific email or file.” (Gmail Help, checked 30 September 2026.) Update the document, not just the draft.
- Tone and format: custom instructions in the Gemini side panel in Gmail keep your preferences “without needing to repeat them in every conversation” (Google Workspace Updates, 2 September 2026).
- Recurring answers: save them as Gmail templates, and adapt the opening for each customer.
And in Outlook?
Draft with Copilot works from your prompt and has tone and length settings (Microsoft Support). Keep a facts document, paste the relevant lines into your prompt, and keep your standing preferences in a short text you reuse. The correction lives in your documents rather than in the tool.
How can you tell whether a tool really learns from you?
Test it, because “learns from your edits” appears on many product pages.
- Pick one clear correction, such as a refund window or a sign-off.
- Make it on three drafts in a row, the same way each time.
- Watch the next five drafts on the same kind of question.
- Ask the tool what it learned, if it can show you. A written, readable record is a good sign.
- Check it stays fixed after a week.
If the same mistake comes back, the correction is not being stored. Move it to a place that sticks: the source or a standing instruction.
What habits make corrections work?
- Correct consistently. Two people fixing the same thing in two different ways teaches a tool nothing clear.
- Fix facts and tone separately. A single rewrite that changes both is hard for any tool to interpret, and hard for you to review later.
- Keep a short list of recurring corrections. If the same item appears three times, it belongs in a source or an instruction.
- Review the list monthly, and delete what no longer applies.
How does ReplyPen handle corrections?
When you edit a Draft before sending, ReplyPen notices the difference and adjusts. You can also give Feedback on a Draft (“too long”, “we don’t offer that”) and get a new one. What it learns is written in plain language in its Knowledge log and Tone profile, in Memory, where you can read it and correct it. And because each Draft’s Check note shows its sources, you can see when a wrong fact came from an outdated page, and fix the page.
For more on how tools learn your style, read What “learns your tone” actually means in an AI email tool. For when Gmail’s own AI is enough, see Is the free AI in Gmail and Outlook enough for customer email?.
