Inbox-Triage — Recurring Email Triage

Paired with inbox-setup. This skill consumes the 7-file knowledge base that inbox-setup writes at ${WORKSPACE}/Email/. The file contracts MUST match exactly. See references/kb_file_contract.md — this is the mirror of the setup-side contract, viewed from the read side.

Run on a recurring schedule (1–3x daily) or on demand. Classify recent emails, research new senders, generate decision recommendations, draft replies (NEVER SEND), deliver a clean report, and update the knowledge base with what was learned this run.

Invocation Triggers

• "triage my inbox"
• "inbox triage"
• "check my email"
• "run email triage"
• "process my inbox"
• "what's new in my email"
• "handle my email"
• "email triage"

Prerequisites

Required reads at start (fail-fast if missing):

Core (required):

• ${WORKSPACE}/Email/email-taxonomy.md — classification + report preferences
• ${WORKSPACE}/Email/email-patterns.md — voice, persona, templates, hard rules

Optional core (read if exists):

• ${WORKSPACE}/Email/evaluation-framework.md
• ${WORKSPACE}/Email/rate-card.md

Evolving (read AND update every run):

• ${WORKSPACE}/Email/blocklist.md
• ${WORKSPACE}/Email/tracker.md

Output:

• ${WORKSPACE}/Email/triage-log/<YYYY-MM-DD>-<run-label>.md — per-run log

If any core required file is missing → halt, direct user to run inbox-setup first. Use scripts/kb_reader.py to perform the read + validation.

DRAFTS ONLY — Never Send

This skill creates drafts. It NEVER sends.

This is the safety property that makes the skill safe to run automatically. Stated multiple times in this skill body. Non-negotiable.

The scripts/draft_safety_validator.py enforces it post-run. Any send-shaped tool call in the action log fails validation. See references/drafts_only_safety.md for the full discipline canon.

Step 0: Grill-Me Intake (Light — 0–2 Optional Override Questions)

Inbox-triage is light-intake by design — it runs on a recurring cadence with preferences pre-baked into the knowledge base from inbox-setup. The grill-me discipline here is asking ONLY the override questions that matter THIS run.

Q1 (optional, asked only when on-demand run is outside normal cadence)

Override the default 9-hour search window? Pick: yes (specify hours) / no (use default).

Why I'm asking: If you're running on-demand outside your normal 2x/day cadence, you may want a wider window (24h after a long break) or narrower (2h for a quick check).

Skip if cadence is normal.

Q2 (optional, asked only when user invokes with category-skip intent)

Skip any categories this run? E.g., "skip newsletters", "skip financial".

Why I'm asking: Sometimes you just want to scan opportunities or just want to clear active threads. Category skip narrows the run scope.

Skip if user gave no category-skip signal.

Stop condition: Max 2 questions. Default invocations skip both questions and run with KB-default preferences. The skill is optimized for fast recurring execution; intake is the exception, not the norm.

Step 1: Determine Search Window

Compute via current date math. Default lookback: 9 hours (works for 2x/day cadence with slight overlap so emails between runs aren't missed).

Use scripts/search_window_calculator.py --cadence <CADENCE> --now <ISO>:

now = current_datetime
window_start = now - 9_hours   (default for 2x-daily)
run_label = "Morning" if now.hour < 12 else "Afternoon" if now.hour < 17 else "Evening"

Cadence-to-default-window mapping (override via Q1):

Step 2: Email Search

Two queries (provider-agnostic adapter pattern):

• Primary: Inbox + sent after window_start
• Secondary: Starred unread (catch flagged items missed in primary)

Collect for each email: sender, subject, date, snippet, thread ID, labels.

Provider adapter mapping:

Step 3: Classification

Apply the taxonomy from email-taxonomy.md. For lowest-priority category (newsletters / automation / spam): skip thread reads entirely — context cost not worth it. For everything else: read full thread.

Step 4: Sender Research

For senders not in tracker / blocklist / prior logs:

1. Check blocklist.md → if matched, auto-skip
2. Check tracker.md → if known thread, note existing context
3. For opportunity senders (per evaluation framework): web search for company legitimacy, social presence, intermediary status

Skip research entirely for: known senders (in tracker), internal email, automated notifications, obvious low-priority.

Step 5: Recommendations

For decision-required emails, apply the framework from evaluation-framework.md. Categorize:

Each: brief "why", relevant context, pricing/timeline comparison if applicable.

Skip Step 5 entirely if no evaluation-framework.md exists.

See references/triage_decision_framework.md for the framework canon.

Step 6: Drafts

For every reasonable reply candidate, create a draft using email-patterns.md voice rules.

Draft for: opportunity responses (TAKE IT / WORTH / PASS), active conversations needing reply, action items, important personal emails.

Do NOT draft for:

• Clearly no-response emails (newsletters, automation, FYI)
• Threads where user already replied
• Blocked senders (unless new info changes the calculus)

Mechanics:

• Draft only in the existing thread when possible (preserves context)
• Set to, subject (Re: [original])
• NEVER call any send operation. Only create drafts.

The draft body MUST honor:

• Voice register from email-patterns.md
• Forbidden tokens (S3.Q2 pet peeves)
• Sign-off patterns
• Persona context
• Hard rules (S3.Q6 — non-negotiable)
• Reply length per email-patterns.md

If evaluation-framework.md exists, draft tone matches recommendation:

• TAKE IT → engaged + concrete next step
• WORTH → curious + 1-2 clarifying questions
• PASS → polite decline + brief reason (no hedging promises)
• FLAG → NO draft

Step 7: Report Delivery

Honor user's preference from email-taxonomy.md "Report Preferences" section. Default: email draft to self with HTML.

Subject: Inbox Triage — [Day], [Month Date] ([Run Label])

Sections (in order):

1. Overview — 2–3 sentences. What happened? Anything urgent?
2. Stats — Counts: processed, drafts created, action needed, skipped.
3. Action Needed — Overdue items, decisions, drafts to review, deadlines.
4. Quick Reference — One line per email, alphabetical by sender. **Sender** — one-sentence summary + recommendation.
5. Detailed Cards — Opportunities, active threads, flags. Each: sender / subject / category, recommendation + reasoning, key context. NO draft text previews (drafts are already in email client for user to read there).
6. Footer — Generation timestamp + KB update summary.

Formatting (if HTML):

• Inline CSS only (Gmail strips <style>)
• Color-coded by recommendation:
  • green → TAKE IT
  • am