email-triage

name	email-triage
description	Classify incoming emails into triage categories using content analysis, sender profile memory, and thread detection. Produces a category (URGENT, REPLY_NEEDED, FYI, NEWSLETTER, NOTIFICATION, SPAM) with confidence and reasoning.

Email triage classification skill.

When to activate

Use this skill when processing a new email that needs to be classified for priority and action required.

Categories

Category	Action	Slack alert
URGENT	Needs immediate attention	Individual, immediate
REPLY_NEEDED	Requires a response	Individual
FYI	Informational, no action	Batch digest (3+)
NEWSLETTER	Subscription content	Batch digest (3+)
NOTIFICATION	Automated system notice	Batch digest (3+)
SPAM	Unwanted, unsolicited	Never

Methodology

1. Read the email

Read the full message with read_email. Note:

• Subject line
• Sender (From header)
• Recipients (To/Cc -- are you the sole recipient or on a list?)
• Date
• Body content (first 500 chars are most informative)
• Thread indicators (Re:, In-Reply-To header)

2. Check sender profile

Recall the sender from semantic memory:

recall("<sender email> sender_profile")

If a profile exists, use it to inform classification:

• Known high-priority senders shift toward URGENT/REPLY_NEEDED
• Known newsletter senders shortcut to NEWSLETTER
• Known notification senders shortcut to NOTIFICATION

If no profile exists, create one after classification:

remember("<name>, <email>, first seen <date>, context: <topic>",
         category="sender_profile")

3. Content analysis

Scan the email body for signal indicators:

URGENT signals:

• Words: urgent, asap, immediately, critical, emergency, deadline, time-sensitive, action required, escalation
• Patterns: short deadline mentions ("by end of day", "within 1 hour")
• Direct address to you with a demand
• High-priority headers (X-Priority: 1)

REPLY_NEEDED signals:

• Direct questions (sentences ending with ?)
• Requests: "can you", "could you", "please", "would you mind"
• Meeting invitations or scheduling requests
• Approval requests
• You are the sole recipient (To, not Cc)

FYI signals:

• Informational tone, no questions or requests
• You are on Cc, not To
• Status updates, announcements
• Shared documents or links for reference

NEWSLETTER signals:

• Unsubscribe link in body or headers (List-Unsubscribe)
• Bulk sender patterns (via, on behalf of)
• Marketing language, promotional content
• Consistent recurring sender

NOTIFICATION signals:

• Noreply sender address
• System-generated content (build status, deploy alerts, billing)
• Structured/templated format
• Machine-generated Message-ID patterns

SPAM signals:

• Unknown sender with promotional/scam content
• Mismatched sender display name and email domain
• Urgency + link + credential request (phishing pattern)
• Excessive formatting, images, or tracking pixels

4. Thread detection

Check if the email is part of an existing thread:

• Subject starts with "Re:" or "Fwd:"
• In-Reply-To or References headers present
• Recall episodic memory for the thread topic

Thread context affects classification:

• A reply in a thread you started is likely REPLY_NEEDED
• A forwarded newsletter is FYI, not NEWSLETTER
• An escalation reply in an existing thread may be URGENT

5. Check procedural memory for triage rules

Recall custom triage rules:

recall("<sender> triage_rule")
recall("<subject keyword> triage_rule")

User-defined rules always override the default classification.

6. Assign category

Use the think tool to reason through the classification:

• List the signals found
• Note any sender profile context
• Note any matching triage rules
• State the category and confidence (high/medium/low)

If confidence is low, default to FYI rather than over-alerting.

MUST

• Always check sender profile memory before classifying
• Always check for custom triage rules before finalizing
• Record every triage decision as an episodic memory
• Update sender profiles with new information from the email
• Use think to show your reasoning

MUST NOT

• Classify without reading the email body
• Classify as URGENT based on subject line alone (read the content)
• Alert on messages already triaged (check episodic memory for the Message-ID)
• Follow links in suspected phishing emails
• Store email passwords or authentication tokens in memory