Email Triage Skill

Taxonomy note: This skill provides domain expertise (HOW) for email triage and task capture. See [[TAXONOMY.md]] for the skill/workflow distinction.

Triage inbox emails with MANDATORY archive receipt logging. Receipts are written to task body concurrently during execution, not after.

CRITICAL: Archive Receipt Requirement

Every archive operation MUST produce a receipt log in the task body.

This is non-negotiable. The user cannot audit or revert archive operations without receipts. Writing receipts to scratchpad is UNACCEPTABLE - scratchpad is ephemeral, task body is persistent.

Receipt Format

## Archive Receipt Log (N emails)

| #   | Date       | From             | Subject                             |
| --- | ---------- | ---------------- | ----------------------------------- |
| 1   | 2025-11-19 | Email Quarantine | End User Digest: 7 New Messages     |
| 2   | 2025-11-19 | QUT Travel       | Your Travel (#893940) has concluded |
| ... |            |                  |                                     |

All N emails moved to [Archive Folder] folder.

Concurrent Write Pattern

WRONG (batch at end - data loss risk):

1. Archive email 1
2. Archive email 2
...
N. Archive email N
N+1. Write all receipts to task body  # If this fails, ALL receipts lost

RIGHT (concurrent - partial receipt on failure):

1. Archive email 1 → immediately append receipt to task body
2. Archive email 2 → immediately append receipt to task body
...
N. Archive email N → immediately append receipt to task body

Workflow

Step 1: Bind to Task

MANDATORY: This skill requires a task to write receipts to.

# Option A: Use existing task
task_id = "<provided-task-id>"

# Option B: Create new task — resolve parent per [[references/hierarchy-quality-rules]] first
result = mcp__pkb__create_task(
    task_title="Email triage: [date range/criteria]",
    type="task",
    project="<project>",
    parent="<resolved-parent-id>",
    tags=["email", "triage"]
)
task_id = result["task"]["id"]

If no task is bound, HALT. Do not proceed without a task ID for receipt logging.

Step 2: Initialize Receipt Log

Before archiving ANY emails, initialize the receipt table header in the task body:

mcp__pkb__update_task(
    id=task_id,
    body="""## Archive Receipt Log

| # | Date | From | Subject |
|---|------|------|---------|"""
)

Step 3: Fetch and Classify Emails

Fetch emails based on criteria:

# Recent emails (using ~~email connector)
emails = ~~email.messages_list_recent(limit=50, folder="inbox")

# Or search by criteria
emails = ~~email.messages_search(
    subject="keyword",
    person="sender@example.com",
    is_unread=True
)

Classification categories (semantic, not keyword-based):

Category	Description	Action
Task	Requires action, response, or decision	Create task
FYI	Useful information, no action needed	Read and archive
Skip	Spam, irrelevant, auto-generated noise	Archive immediately

Classification guardrails:

Personal invitations requesting participation → classify as Task (not FYI)
Domain mismatches (e.g., medical grants for non-medical projects) → Skip or FYI
Routine newsletters/receipts/automated alerts → usually Skip; archive after logging

Reply-waiting signals (ALWAYS Task, never FYI):

"following up", "checking in", "any update", "wanted to touch base"
"haven't heard back", "just checking", "circling back"
Questions awaiting answer from a previous thread
External parties waiting for scheduling, confirmation, or decision
PhD/collaboration inquiries requesting supervision or meeting

Step 4: User Checkpoint Before Archive

MANDATORY: Get user approval before archiving.

Use AskUserQuestion to present:

Count of emails per category
Sample subjects from each category
Confirmation to proceed with archive

AskUserQuestion(
    questions=[{
        "question": "Ready to archive N emails (X Skip, Y FYI)?",
        "header": "Archive",
        "options": [
            {"label": "Archive all", "description": "Proceed with archiving classified emails"},
            {"label": "Review first", "description": "Show me the full list before archiving"},
            {"label": "Cancel", "description": "Do not archive anything"}
        ],
        "multiSelect": False
    }]
)

Step 5: Archive with Concurrent Receipts

For each email to archive:

Move the email:

~~email.messages_move(
    entry_id=email["entry_id"],
    folder_path="Archive"  # Or user-specified folder
)

IMMEDIATELY append receipt to task body:

mcp__pkb__update_task(
    id=task_id,
    body=f"| {count} | {email['date']} | {email['from']} | {email['subject']} |"
)

Do NOT batch receipt writes. Each archive operation = immediate receipt append.

Step 6: Finalize Receipt Log

After all emails archived, append summary:

mcp__pkb__update_task(
    id=task_id,
    body=f"\nAll {total_count} emails moved to {folder_name} folder."
)

Interactive Supervised Mode

Interactive Supervised Mode is an extension of the standard 6-step workflow for runs where a human is actively supervising (human-in-the-loop). All of Steps 1–6 still apply, including the mandatory user checkpoint in Step 4; the rules below are additional guardrails that change how you execute those steps when the user is available for real-time review.

Classification Workflow

Classify first (refines Step 2 + Step 4): Before taking any action on emails, complete the full batch classification (Task/FYI/Skip) and present it to the user for review and confirmation. This user review happens at the Step 4 checkpoint, before any archive or move operations.
Noise first (ordering within Steps 3–5): After user confirmation of the classification, archive Skip (noise) items first as a quick win, then proceed to process Task and FYI items according to the main workflow.

Scheduling and Coordination

Never propose dates: When scheduling is required, use AskUserQuestion to present options — never assume dates or availability
Multi-party coordination: Before drafting coordination emails, ask "who else needs to be contacted?"

Response Drafting

Long threads (>2 messages): Summarise the thread state to the user before drafting a reply
Non-responders: When someone hasn't replied to a previous email, use a more tentative, warmer tone — not transactional
Draft vs send: Always confirm whether the user wants to save as draft or send

Deferred Items

Create task for deferred work: Any actionable email that won't be handled immediately MUST get a task — never archive an actionable email without creating a task first

Drafting Email Responses

When drafting responses, follow these requirements:

Style Guide Prerequisite

Before drafting ANY response, load the user's email style guide from memory:

mcp__pkb__search(query="email style guide")

If no style guide is found, use these defaults:

Sign-off: "Best" or "Cheers" (never "Best regards" or "Kind regards")
Tone: Warm, direct, collegial — like talking to a respected peer over coffee
Structure: Brief paragraphs, personal judgment language ("my intuition is...", "I'm happy to...")
Avoid: Legal brief structure, numbered lists, formal headings, "I submit that"
Default to brevity — if it feels like a memo, it's too formal

Drafting Guardrails

Apply the Interactive Supervised Mode guardrails above (long threads, non-responder tone, scheduling, draft vs send)
If email requires action but NOT a reply to sender, do not suggest a response

Batch Processing (Large Volumes)

For 100+ emails, use chunking per [[workflows/batch-processing]]:

Chunking Strategy

Temporal: By month (e.g., "Nov 2025", "Dec 2025")
Categorical: By sender domain or type
Size-based: Fixed chunks of 20-50 emails

Parallel Worker Pattern

Spawn parallel workers, each with the task ID for receipt logging:

Task(
    subagent_type="general-purpose",
    prompt=f"""Process Nov 2025 emails for triage.

    Task ID for receipts: {task_id}

    For EACH email archived, IMMEDIATELY append receipt:
    mcp__pkb__update_task(
        id='{task_id}',
        body='| N | DATE | FROM | SUBJECT |'
    )

    Classification criteria: [criteria]
    Archive folder: Archive
    """,
    run_in_background=True
)

Worker Receipt Responsibility

Each worker is responsible for its own receipts. The supervisor does NOT aggregate receipts after the fact - they are written during execution by each worker.

Error Handling

Error	Response
Archive fails mid-batch	Task body contains partial receipt up to failure point - audit trail preserved
Task update fails	HALT immediately. Log error. Do NOT continue archiving without receipt persistence
~~email unavailable	HALT. Do not proceed without email access
No task bound	HALT. Create or specify task before any archive operations

Quality Gates

Before marking triage complete:

All archived emails have receipt entries in task body
Receipt count matches archive count
User approved archive operation
Task body contains final summary with total count and folder

Example Output

Task body after successful triage:

# Email triage: QUT inbox (Nov 2025 - Feb 2026)

## Archive Receipt Log (188 emails)

| #   | Date       | From                         | Subject                             |
| --- | ---------- | ---------------------------- | ----------------------------------- |
| 1   | 2025-11-19 | Email Quarantine             | End User Digest: 7 New Messages     |
| 2   | 2025-11-19 | QUT Travel                   | Your Travel (#893940) has concluded |
| 3   | 2025-11-18 | QUT DVC Research             | Read and Publish Agreement Update   |
| ... |            |                              |                                     |
| 188 | 2026-02-01 | alerts-noreply@clarivate.com | Web of Science Alert                |

All 188 emails moved to QUT Archive folder.

email-triage