AI Self-Improvement Tutorial: How Your AI Learns Your Habits

You don't configure this feature — you just use the app, and it studies you. Every draft you reject, every edit you make before sending, every rewrite instruction becomes a signal. As feedback accumulates, the AI analyzes it and works out what you keep correcting. Improvements the evidence clearly supports apply automatically, and every one can be revoked.

This tutorial shows you where to check what changed and how to make it learn faster.

🎬 2-minute demo path — showing this live?

1. Open Dashboard → Business Intelligence → Learning history → read one observation out loud: the AI noticed a pattern in your corrections.
2. Open an entry under Applied improvements → tap View change to show exactly what it added, then point at Revoke.
3. Point at an Auto-tuned Settings row: "it noticed I kept rejecting day-2 follow-ups, so it spaced them out itself."

---

Step 1 — See what it has noticed

Open Dashboard, switch to the Business Intelligence tab, and choose Learning history. The top section, Learning observations, lists patterns found in your recent rejections, edits, and rewrite instructions. Each entry shows its type, its status, and its Evidence — the actual drafts it learned from.

Observations come in three kinds:

• Style — "Replies to price questions should lead with the price, not the pleasantries." When the same correction shows up in at least two separate drafts you edited or rewrote, and it doesn't conflict with your existing reply preferences, it is added to them automatically.
• Data gap — the AI keeps getting a fact wrong. If you typed the correct value yourself and it belongs to a basic company field, such as lead time, MOQ, payment terms, or website, the company data is updated. Anything else stays an observation for you to fix in company info — facts belong in your profile, not in the AI's habits.
• Insight — something worth knowing, such as a pattern in what customers push back on. Insights are never applied on their own.

The status on each entry tells you what happened: Observing, Applied, No change needed, Not applied, or Undone.

✓ Done when you've read your first observations and opened the evidence behind one of them.

---

Step 2 — Review and revoke what it applied

Below sits Applied improvements — everything currently in force. Three things to look at:

• View change shows the before and after, so you can see the exact rule or field that changed.
• The effect line on each improvement: "This issue: 5 in the 7 days before → 1 since accepting." That's the improvement proving (or failing to prove) its worth.
• Revoke removes a change that stops helping. If you've edited that content since, your later edits are kept rather than overwritten with older text.

Follow-up pacing has its own list, Auto-tuned Settings. If you keep rejecting day-2 follow-up drafts, the AI concludes you find them pushy and stretches the interval itself ("Follow-up interval: 2 → 4 days"). It's only ever allowed to make follow-ups less frequent, never more — and Revoke restores the old pace.

✓ Done when you've opened one applied improvement, read its change, and revoked anything that no longer fits.

---

Step 3 — Teach it faster (this is the real skill)

The AI learns from labeled corrections much faster than from silent ones. Three habits multiply its learning speed:

1. Pick a reason instead of a bare reject. Choosing Wrong tone or Wrong fact on a draft tells it exactly what was wrong; a plain reject leaves it guessing. The reject reasons sit under desktop drafts and in Telegram, Slack and WeCom notifications.
2. Type the instruction when you have 5 seconds. Use Tell AI how to adjust on a draft: "Include shipping in the price" is a premium learning signal — specific, actionable, in your own words.
3. Edit rather than rewrite from scratch. Your before/after diff is studied — approving an edited draft shows it precisely what you changed.

Analysis runs in the background as new feedback accumulates. Impatient? There's an Analyze now button in the panel.

---

Things that trip people up

"No observations yet." It needs material — a handful of rejects, edits, or rewrites before the first analysis. A brand-new install has nothing to learn from; a week of Suggest-mode usage does.

"It applied something wrong." Revoke it. And if the root cause is a wrong fact, fix it in company info — facts always beat habits.

"Will it start doing things I didn't approve?" It can add a style rule backed by repeated corrections, update a basic company field you corrected yourself, and space out follow-ups. Every change is listed with its evidence and can be revoked. It never makes follow-ups more frequent, and learning never gives the AI new sending permissions.

---

Next: Put the trained AI to work on quiet customers →