Templates
Email triage
Decide whether an email needs the recipient soon, and what kind of message it is.
Decide whether an email needs the recipient soon, and what kind of message it is. Pattern: composite_scoring. Model: jev-1.13.0. Outage rule: review.
When to use it
- An inbox or shared mailbox where people lose time sorting mail by hand.
- You already ask an LLM "is this urgent?" and act on a yes or no.
- Urgency means a person wrote it, someone is waiting, and ignoring it costs something.
When not to use it
- You need a reply drafted. That is text generation; keep an LLM for it and use this set to decide which mail gets one.
- The rule is exact, such as a sender domain or a subject tag. Keep that in code or in a spec check.
Questions
| Id | Type | Asks | Answers | Gating |
|---|---|---|---|---|
real_person | noul | Was email written by a real person to the recipient, rather than sent by an automated system, newsletter, or marketing tool? | true or false | yes |
someone_waiting | noul | Is someone the recipient knows waiting on the recipient to reply or act? | true or false | yes |
cost_of_ignoring | score | What will it cost the recipient to ignore email for a week? | 4 levels | yes |
category | choice | What kind of message is email, judged by what the sender wants from the recipient? | work_request, scheduling, personal, newsletter, transactional, none_of_these | no |
work_type | choice | Assume email is a work request to the recipient. What does the sender need the recipient to do? | decision, information, review, none_of_these | yes |
Reading the result
- App code branches on
overallActionfirst. Act only onauto;reviewgoes to a person. - Then read
route:urgent,read_laterornormal. - The
urgencycomposite weighs three checks. Tune the weights andlevelThresholdson your own labeled mail. work_typeis speculative: it only counts whencategoryiswork_request.
Spec
Save it as bandwise/sets/email-triage.json and edit it for your data.
{
"schemaVersion": 1,
"model": "jev-1.13.0",
"input": {
"schema": {
"type": "object",
"required": [
"email",
"me"
],
"properties": {
"email": {
"type": "object",
"required": [
"from",
"subject",
"body"
],
"properties": {
"from": {
"type": "string"
},
"subject": {
"type": "string"
},
"body": {
"type": "string"
},
"signature": {
"type": "string"
}
}
},
"me": {
"type": "object",
"properties": {
"name": {
"type": "string"
},
"known_contacts": {
"type": "array",
"items": {
"type": "string"
}
}
}
}
}
},
"redactPaths": [
"email.signature"
]
},
"stages": [
{
"id": "triage",
"questions": {
"real_person": {
"type": "noul",
"instructions": "Was `email` written by a real person to the recipient, rather than sent by an automated system, newsletter, or marketing tool?",
"criteria": {
"true": "A person wrote this message to the recipient.",
"false": "Automated, bulk, newsletter, receipt, or marketing mail."
},
"meta": {
"label": "Real person wrote it"
}
},
"someone_waiting": {
"type": "noul",
"instructions": {
"question": "Is someone the recipient knows waiting on the recipient to reply or act?",
"recipient": "`me.name`",
"known_contacts": "`me.known_contacts`",
"email": "`email`"
},
"criteria": {
"true": "The sender or someone named is blocked on, or explicitly waiting for, the recipient.",
"false": "Nobody is waiting on the recipient."
},
"meta": {
"label": "Someone is waiting"
}
},
"cost_of_ignoring": {
"type": "score",
"instructions": "What will it cost the recipient to ignore `email` for a week?",
"criteria": [
"Nothing. Safe to ignore.",
"Minor. A small delay or a mildly annoyed contact.",
"Real. A missed deadline, lost money, or a damaged relationship.",
"Severe. Legal, financial, or safety consequences."
],
"meta": {
"label": "Cost of ignoring"
}
},
"category": {
"type": "choice",
"instructions": "What kind of message is `email`, judged by what the sender wants from the recipient?",
"criteria": {
"work_request": "Someone needs work, a decision, or information from the recipient.",
"scheduling": "Meeting times, invites, or calendar changes.",
"personal": "Friends or family.",
"newsletter": "Subscribed content or digests.",
"transactional": "Receipts, notifications, password resets.",
"none_of_these": null
},
"meta": {
"label": "Category"
}
},
"work_type": {
"type": "choice",
"instructions": "Assume `email` is a work request to the recipient. What does the sender need the recipient to do?",
"criteria": {
"decision": "Make or approve a decision.",
"information": "Send facts, an answer, or a document.",
"review": "Review or give feedback on something the sender made.",
"none_of_these": null
},
"meta": {
"label": "Work type",
"description": "Speculative: only relevant when category is work_request."
}
}
}
}
],
"policies": {
"real_person": {
"type": "noul",
"gating": true,
"noul": {
"trueAt": 0.85,
"falseAt": 0.15,
"reviewMargin": 0.1
},
"actions": {
"high": {
"kind": "auto"
},
"medium": {
"kind": "auto"
},
"low": {
"kind": "review"
}
}
},
"someone_waiting": {
"type": "noul",
"gating": true,
"noul": {
"trueAt": 0.85,
"falseAt": 0.15,
"reviewMargin": 0.1
},
"actions": {
"high": {
"kind": "auto"
},
"medium": {
"kind": "review"
},
"low": {
"kind": "review"
}
}
},
"cost_of_ignoring": {
"type": "score",
"gating": true,
"thresholds": {
"high": 0.7,
"medium": 0.45
},
"actions": {
"high": {
"kind": "auto"
},
"medium": {
"kind": "review"
},
"low": {
"kind": "review"
}
}
},
"category": {
"type": "choice",
"gating": false,
"thresholds": {
"high": 0.6,
"medium": 0.3
},
"actions": {
"high": {
"kind": "auto"
},
"medium": {
"kind": "auto"
},
"low": {
"kind": "fallback",
"config": {
"kind": "value",
"value": "none_of_these"
}
}
}
},
"work_type": {
"type": "choice",
"gating": true,
"relevantWhen": {
"q": "category",
"eq": "work_request"
},
"thresholds": {
"high": 0.65,
"medium": 0.4
},
"actions": {
"high": {
"kind": "auto"
},
"medium": {
"kind": "review"
},
"low": {
"kind": "review"
}
}
}
},
"composites": [
{
"id": "urgency",
"kind": "weighted",
"terms": [
{
"q": "real_person",
"weight": 0.25
},
{
"q": "someone_waiting",
"weight": 0.35
},
{
"q": "cost_of_ignoring",
"weight": 0.4
}
],
"policy": {
"type": "composite",
"gating": true,
"levelThresholds": {
"high": 0.7,
"medium": 0.4
},
"actions": {
"high": {
"kind": "auto"
},
"medium": {
"kind": "review"
},
"low": {
"kind": "auto"
}
}
}
}
],
"routes": [
{
"when": {
"composite": "urgency",
"gte": 0.7
},
"output": "urgent"
},
{
"when": {
"q": "category",
"eq": "newsletter"
},
"output": "read_later"
}
],
"defaultRoute": "normal",
"savings": {
"comparatorModel": "claude-haiku-4-5",
"estOutputTokensPerQuestion": 60,
"kind": "decision"
},
"onUnavailable": "review"
}Example states
The expected outcome is what a person would decide. It is not a recorded model answer.
Contract approval from a known contact
Expected: Urgent: a person wrote it, they are waiting, and a missed deadline costs money.
{
"email": {
"from": "ana@example.com",
"subject": "Re: contract renewal",
"body": "Hi Sam, legal signed off on the renewal terms. I need your approval on the final contract by Friday so we can keep the current pricing. Can you confirm today?",
"signature": "Ana, Procurement"
},
"me": {
"name": "Sam",
"known_contacts": [
"ana@example.com",
"legal@example.com",
"lee@example.org"
]
}
}Weekly newsletter
Expected: Read later: automated mail, nobody waiting, nothing lost by ignoring it.
{
"email": {
"from": "digest@news.example.net",
"subject": "This week in data engineering",
"body": "Five articles we liked this week, plus a new podcast episode. You are receiving this because you subscribed."
},
"me": {
"name": "Sam",
"known_contacts": [
"ana@example.com",
"legal@example.com",
"lee@example.org"
]
}
}Meeting move from a colleague
Expected: Normal: a person wrote it about scheduling, and a week's delay costs little.
{
"email": {
"from": "lee@example.org",
"subject": "Moving our 1:1",
"body": "Hey Sam, can we push Thursday's 1:1 to next week? No rush, whenever suits you."
},
"me": {
"name": "Sam",
"known_contacts": [
"ana@example.com",
"legal@example.com",
"lee@example.org"
]
}
}Borderline cases
One case near the line for each question. Use them to test your wording before you trust the thresholds.
real_person
A sales sequence tool sent it, but it is written in the first person and uses the recipient's name.
{
"email": {
"from": "jordan@vendor.example.com",
"subject": "Quick question, Sam",
"body": "Hi Sam, I noticed your team is hiring data engineers. Would a 15 minute call next week make sense? Jordan"
},
"me": {
"name": "Sam",
"known_contacts": [
"ana@example.com",
"legal@example.com",
"lee@example.org"
]
}
}someone_waiting
An FYI that ends with an open invitation, not a request.
{
"email": {
"from": "lee@example.org",
"subject": "Notes from the planning session",
"body": "Sharing my notes from today. Let me know if you have thoughts."
},
"me": {
"name": "Sam",
"known_contacts": [
"ana@example.com",
"legal@example.com",
"lee@example.org"
]
}
}cost_of_ignoring
A payment reminder with a due date weeks away: minor now, real later.
{
"email": {
"from": "billing@saas.example.com",
"subject": "Your invoice is due in 3 weeks",
"body": "Invoice 2291 for 480 USD is due at the end of next month. Late payments pause the account."
},
"me": {
"name": "Sam",
"known_contacts": [
"ana@example.com",
"legal@example.com",
"lee@example.org"
]
}
}category
A meeting invite that also asks for a decision before the meeting.
{
"email": {
"from": "ana@example.com",
"subject": "Vendor review Tuesday",
"body": "Sending an invite for Tuesday. Before then, please pick which of the two vendors we should drop."
},
"me": {
"name": "Sam",
"known_contacts": [
"ana@example.com",
"legal@example.com",
"lee@example.org"
]
}
}work_type
"Take a look and sign off" is both a review and an approval.
{
"email": {
"from": "legal@example.com",
"subject": "NDA for the partner pilot",
"body": "Attached is the NDA. Can you take a look and sign off so we can send it today?"
},
"me": {
"name": "Sam",
"known_contacts": [
"ana@example.com",
"legal@example.com",
"lee@example.org"
]
}
}Try it
Save an example state as state.json, then run the spec locally. Local mode makes no network call and needs no key; answers are synthetic unless a recorded fixture matches.
pnpm bandwise run --local bandwise/sets/email-triage.json state.jsonBandwise is an independent product built on TypeSafe's System One models. It is not TypeSafe's documentation. For the System One models themselves, see docs.typesafe.ai.