Getting to near-perfect sorting
Viltreon works out of the box, but five minutes of setup turns “mostly right” into “almost always right.” This guide shows you exactly which levers matter and how to use them.
How a sorting decision is made
When an email arrives, the AI sees four things: the sender, the subject, the beginning of the message body, and your setup — the list of labels you let it use, each label’s description, and your sorting rules. It picks exactly one label and reports how confident it is. If confidence is below 60%, the email goes to your fallback label instead of being guessed into the wrong place.
That means accuracy is mostly about one thing: how clearly your labels communicate what belongs in them. The AI can only be as precise as the categories you give it.
1. Start with label names that don’t overlap
Before touching descriptions, look at your label list the way a new assistant would. If two labels could plausibly claim the same email, the AI will sometimes pick the one you didn’t want. Common trouble pairs:
- “Finance” and “Receipts”— where does a bank statement go? An Amazon order confirmation? Either merge them or draw the line in their descriptions.
- “Work” and “Projects”— nearly everything work-related is also project-related. Be specific about the split.
- “Misc”, “Other”, “Stuff”— keep at most one catch-all, and make it your fallback label.
Nested labels work well and the AI sees the full path (Work/External, Work/Internal), which itself carries meaning.
2. Descriptions: the highest-leverage tool you have
Every label can carry a description (the document icon next to a label on the Inbox Rules page). The AI reads these on every single decision. A good description does two jobs: it says what belongs in the label, and it draws the boundary against the labels it could be confused with.
Weak description:
Finance stuff
Strong description:
Receipts, invoices, order and payment confirmations, bank and credit card statements, tax documents. NOT promotional emails from stores — those go to Shopping.
Three patterns that work:
- List concrete examples.“Order confirmations, shipping updates, return labels” beats “purchase-related email.” The AI matches incoming mail against your examples.
- Name senders when they define the category.“Anything from my accountant, my bank, Stripe, or PayPal” is unambiguous.
- Say what does NOT belong.One “NOT X” clause resolves most boundary fights between similar labels. Put it in whichever label keeps getting the wrong mail.
Keep it tight: one or two precise sentences outperform a paragraph.
3. Sorting rules: for logic that crosses labels
Sorting rules live in Settings and apply to every decision. Descriptions define categories; rules define behavior— priorities, tie-breakers, and special cases. The AI is instructed to follow them strictly. Use them for:
- Tie-breakers:“If an email is both a receipt and a shipping update, prefer Finance over Shopping.”
- Sender routing:“Anything from @mycompany.com goes to Work/Internal, no matter the topic.”
- Intent detection:“Cold outreach and sales pitches from people I have never emailed go to Pitch.”
- Personal exceptions:“Emails from my mom (jane@example.com) always go to Family, even if they look like forwards or newsletters.”
Tips for writing them:
- Write each rule as one plain sentence: if this, then that label. Bullet-like lines work well; clever prose does not add anything.
- Refer to labels by their exact names so there is no ambiguity about which label you mean.
- Keep rules for exceptions and priorities. If you find yourself describing what a label IS, that sentence belongs in the label’s description instead.
- Fewer, sharper rules beat many vague ones.
4. Hide labels the AI should never use
The eye icon next to each label on the Inbox Rules page controls whether the AI can file mail into it. Hide labels you manage by hand (“To Print”, “Waiting on reply”) and any label that exists for archiving rather than incoming mail. Every label you hide is one less way to be wrong — a shorter menu makes every remaining choice more accurate.
5. Choose a fallback label you actually review
When the AI is less than 60% sure, it refuses to guess and files the email under your fallback label. This is a feature: uncertain mail is concentrated in one place instead of being scattered into wrong categories. Pick something like “Other” and skim it occasionally — every email you find there is telling you which description or rule to sharpen.
The improvement loop
When an email lands in the wrong place, fix the cause, not the email:
- Wrong label won?Add a “NOT ...” clause to the label that wrongly claimed it, or a tie-breaker rule naming both labels.
- Landed in fallback?The right label’s description was too vague to clear the confidence bar — add the missed email’s type or sender to it as an example.
- A whole sender keeps missing? One sender-routing rule in Settings ends it permanently.
Two or three rounds of this in your first week typically gets sorting to the point where you stop thinking about it — which is the whole idea.