📬 Inbox Intelligence

A communications intelligence skill that turns your overflowing inbox into three things:

1. An action list — emails/messages you actually need to respond to, ranked by urgency
2. A knowledge compendium — everything you should know, organized by topic, cross-referenced, searchable by any agent
3. Peace of mind — the 280 emails that were noise are gone; the 20 that matter are visible

This is not an email summarizer. It's a knowledge extraction and triage system that treats your inbox as an unstructured data stream, applies entropy-aware processing to separate signal from noise, and builds a persistent, growing intelligence layer that makes you smarter over time.

Design target: 300 unread → fully triaged in under 1 hour.

🧠 Core Philosophy

• Your inbox is not a task list — it's an unstructured data stream. Treat it like one.
• Triage is not reading — triage is classification. Reading happens later, only for what matters.
• The compendium outlives the email — an email gets archived; the knowledge it contained stays findable forever.
• Links are first-class citizens — every compendium entry links back to the source message and preserves attachment links.
• The system gets smarter — topic connections, prediction patterns, and cross-references improve with every triage session.
• Manual reads count too — emails you've already handled can still be logged to the compendium. Knowledge capture is independent of triage status.
• Never make me think about where info is — if it came through email or Teams, the compendium should have it. Ask the agent, get the answer.

🏗️ ANCT Architecture

Entropy Profile

Phase:      1          2          3          4          5          6          7
Entropy:    E1         E2         E3         E4         E3         E2         E4→E3
Mode:       DELEGATE   DELEGATE   NARRATE    GEN→NAR    NARRATE    DELEGATE   GEN→NAR

            pull       classify   triage     extract    build      produce    learn &
            messages   by type    & rank     knowledge  compendium action     connect
                                                        entries    list

Why This Entropy Map

📥 Intake: Customize This Skill

{{MESSAGE_SOURCES}}

{{COMPENDIUM_PATH}}

./COMPENDIUM.md

{{LOOKBACK_WINDOW}}

{{KNOWN_PRIORITIES}}

{{NOISE_PATTERNS}}

{{TOPIC_TAXONOMY}}

{{TRIAGE_MODE}}

{{INCLUDE_HANDLED}}

{{ATTACHMENT_HANDLING}}

Phase 1: PULL MESSAGES

Control Mode: DELEGATE | Entropy: E1 (Deterministic)

Pure data retrieval. No judgment.

Actions

1. Connect to

   {{MESSAGE_SOURCES}}

2. Pull all messages within

   {{LOOKBACK_WINDOW}}

3. For each message, capture:

message:
  id: unique-id
  source: outlook | teams | gmail | slack
  type: email | teams-chat | teams-channel | thread-reply
  from: sender name + address
  to: recipients
  cc: cc list
  date: timestamp
  subject: subject line
  body_preview: first 200 chars
  has_attachments: bool
  attachment_names: list
  attachment_links: list  # preserve direct links
  thread_id: conversation thread ID
  is_read: bool
  is_replied: bool
  importance: sender-flagged priority
  channel: Teams channel name (if applicable)

4. If

   {{INCLUDE_HANDLED}}

   is true, pull read messages too (flagged

   is_read: true

   )
5. Deduplicate: collapse thread replies into conversation objects

Exit Condition

All messages pulled, deduplicated, structured. Count logged.

Output

Pulled [N] messages ([X] unread, [Y] read) from [sources] over [window]
Collapsed into [M] conversations

Phase 2: CLASSIFY BY TYPE

Control Mode: DELEGATE → NARRATE (on mismatch) | Entropy: E2 (Procedural)

Pattern-match each message into a category. Same rules every time.

Classification Taxonomy

Classification Rules

IF sender in {{KNOWN_PRIORITIES}} → at minimum 🟡, check for 🔴
IF sender in {{NOISE_PATTERNS}} → ⚪ (skip)
IF you're only in CC and body has no question → 🟢 or ⚪
IF body contains "?" directed at you → 🔴
IF subject contains "FYI" or "no action needed" → 🟡 or 🟢
IF has_attachments and body is short → 📎
IF is thread_reply and you're not addressed → 🔵

🔍 Signal Check (Dynamic Entropy Sensor)

After initial classification, run this check before finalizing any message as

⚪ Noise

🟢 FYI — peripheral

IF sender OR any participant in thread is in {{KNOWN_PRIORITIES}}
   AND subject line is vague (e.g., "Quick question", "FYI", "Heads up", "Checking in")
   THEN:
     → Suspend DELEGATE classification
     → Query the Compendium: "Does this sender's history suggest this
       'Noise' pattern is actually a high-value outlier?"
     → If Compendium shows prior high-value messages with similar vague subjects:
         → Escalate to NARRATE: read the full message body before classifying
     → If no Compendium history: flag for human review with one-line preview

Why this matters: High-priority senders often send critical information under deceptively casual subject lines. The Signal Check prevents over-delegation of messages that look like noise but carry strategic signal.

Exit Condition

Every message has a category. Counts per category logged.

Output

Classification: 🔴 [N] direct requests | 🟡 [N] important FYI | 🟢 [N] peripheral
               ⚪ [N] noise (auto-archived) | 🔵 [N] threads | 📎 [N] document shares
Signal Check: [N] messages escalated from noise/peripheral → NARRATE review

Phase 3: TRIAGE & RANK

Control Mode: NARRATE | Entropy: E3 (Analytical)

This is where judgment begins. Not everything marked 🔴 is equally urgent. Not everything marked 🟢 is equally disposable.

Triage Dimensions

For each non-noise message, score on three dimensions:

Triage Matrix

              High importance     Low importance
High urgency  → RESPOND NOW      → QUICK REPLY
Low urgency   → RESPOND SOON     → LOG TO COMPENDIUM ONLY

Messages with high knowledge value (3+) go to Phase 4 regardless of urgency.

Ranking

Produce a ranked action list:

1. Respond Now (urgency 4-5 + importance 4-5)
2. Respond Soon (urgency 3+ OR importance 3+)
3. Quick Reply (urgency high but importance low — "thanks", "got it", "acknowledged")
4. Read & Archive (no response needed, low knowledge value)
5. Skip (noise — already filtered in Phase 2)

Escalation → GENERATE

If triage is ambiguous (can't determine urgency/importance from preview):

Read the full message body. If still ambiguous, flag for human decision with a one-line summary and ask: "Respond, log, or skip?"

Exit Condition

Every message has urgency + importance + knowledge scores. Ranked action list produced.

Phase 4: EXTRACT KNOWLEDGE

Control Mode: GENERATE → NARRATE (Compression Gate) | Entropy: E4 (Creative)

The highest-value phase. This is where emails become knowledge.

For every message with knowledge value ≥ 2, extract:

Knowledge Extraction Framework

knowledge_entry:
  # === Source ===
  source_id: message-id
  source_type: email | teams-message | teams-channel
  source_link: "deep link to original message"
  source_date: timestamp
  source_from: sender

  # === Content ===
  topic: primary topic (from taxonomy or auto-detected)
  secondary_topics: list of related topics
  summary: 2-3 sentence extraction of what matters
  key_facts: bulleted list of specific facts, numbers, dates, decisions
  quotes: exact quotes worth preserving (with attribution)
  sentiment: positive | neutral | negative | urgent | cautionary

  # === Attachments ===
  attachments:
    - name: filename
      link: "direct link to attachment"
      type: document | spreadsheet | presentation | image | other
      summary: 1-2 sentence description of attachment content (if scanned)
      key_data: extracted data points (if full-extract mode)

  # === Connections ===
  related_entries: list of compendium entries this connects to
  people_involved: who is relevant to this knowledge
  decisions_made: any decisions documented in this message
  action_items: commitments made by anyone (not just you)
  open_questions: unresolved questions mentioned
  deadlines: any dates or timelines mentioned

  # === Meta ===
  confidence: high | medium | low (how sure are we about the extraction)
  compendium_priority: 1-5 (how important is this for long-term reference)

GENERATE Step: Attachment Processing

Based on

{{ATTACHMENT_HANDLING}}

GENERATE Step: Topic Detection

If

{{TOPIC_TAXONOMY}}

1. Scan all extracted summaries for recurring themes
2. Cluster by similarity
3. Propose a taxonomy (user can refine)
4. Assign each entry to 1 primary + 0-3 secondary topics

🔍 Novelty Filter (Pre-Write Gate)

Before writing any entry to the Compendium, run this check:

FOR each knowledge_entry:
  → Search Compendium for matching topic entries from the last 14 days
  → IF entry REPEATS existing facts:
      → Skip creating a new entry
      → Append the new source_link as an additional reference to the existing entry
      → Log: "Duplicate suppressed — source added to [existing entry]"
  → IF entry CONTRADICTS existing facts:
      → TRIGGER GENERATE sub-phase:
          "Fact conflict detected: [old fact] vs [new fact]"
          → Resolve explicitly (e.g., "Deadline shifted from Friday to Tuesday")
          → Update the existing entry with the corrected fact
          → Add a delta note: "Updated [date]: [what changed] — [new source link]"
          → If the change is significant: surface to user in action list as an alert
  → IF entry is NEW (no matching topic or truly novel fact):
      → Proceed to write normally

Why this matters: Without a Novelty Filter, every triage session adds redundant entries and the Compendium grows noisy. The filter ensures the Compendium stays clean, current, and contradiction-free.

⚡ COMPRESSION CHECKPOINT

Every knowledge entry must have:

• A summary ≤ 3 sentences
• At least 1 key fact
• A source link
• A topic assignment

If any field is empty, the entry is incomplete. Re-examine the source.

Escalation → Human

If content is ambiguous or potentially sensitive:

"This message discusses [topic]. Confidence is low on extraction. Should I log it to the compendium? [Yes / Skip / Let me read it first]"

Exit Condition

Knowledge entries produced for all messages with knowledge value ≥ 2. Attachments processed per

{{ATTACHMENT_HANDLING}}

Phase 5: BUILD COMPENDIUM

Control Mode: NARRATE | Entropy: E3 (Analytical)

The compendium is a persistent, growing Markdown knowledge base stored at

{{COMPENDIUM_PATH}}

Compendium Structure

# Communications Compendium
> Last updated: [timestamp]
> Total entries: [N] | Topics: [N] | Sources: [N]

## Table of Contents
- [Topic 1](#topic-1) ([N] entries)
- [Topic 2](#topic-2) ([N] entries)
- ...

---

## Topic 1: [Topic Name]
> [N] entries | Last updated: [date] | Key people: [names]

### [Entry Title] — [Date]
**Source:** [sender] via [email|teams] | [link to original]
**Summary:** [2-3 sentences]
**Key facts:**
- [fact 1]
- [fact 2]
**Attachments:** [filename](link) — [1-line description]
**Decisions:** [if any]
**Open questions:** [if any]
**Related:** [links to related compendium entries]

### [Entry Title] — [Date]
...

---

## Topic 2: [Topic Name]
...

---

## Cross-Reference Index
| Topic | Related Topics | Key People | Recent Activity |
|-------|---------------|------------|-----------------|
| [topic] | [topic, topic] | [names] | [date of last entry] |

---

## People Index
| Person | Topics They Appear In | Last Interaction | Open Items With Them |
|--------|----------------------|------------------|---------------------|
| [name] | [topics] | [date] | [items] |

---

## Timeline
| Date | Topic | Event | Source |
|------|-------|-------|--------|
| [date] | [topic] | [what happened] | [link] |

---

## Decisions Log
| Date | Decision | Made By | Context | Source |
|------|----------|---------|---------|--------|
| [date] | [decision] | [who] | [topic] | [link] |

Compendium Update Rules

1. New entries append under their topic section
2. Existing topics get new entries added chronologically
3. New topics get a new section created
4. Cross-references are updated when related entries are added
5. People index is updated when new people appear in entries
6. Timeline gets a new row for significant events
7. Decisions log captures any decisions mentioned in messages

Deduplication

If a knowledge entry covers the same information as an existing entry:

• Update the existing entry with new details
• Add the new source link as an additional reference
• Don't create a duplicate

Exit Condition

All Phase 4 knowledge entries written to compendium. Indexes updated. Cross-references current.

Phase 6: PRODUCE ACTION LIST

Control Mode: DELEGATE | Entropy: E2 (Procedural)

Structured output from triage decisions. Template work.

Action List Output

# Inbox Triage — [Date]
> Processed: [N] messages | Time: [duration]
> Sources: [email, teams, ...]

## 🔴 Respond Now ([N])
| # | From | Subject | Why It's Urgent | Source |
|---|------|---------|-----------------|--------|
| 1 | [name] | [subject] | [one-line reason] | [link] |
| 2 | ... | ... | ... | ... |

**Suggested responses:** _(for each, a 1-2 sentence draft if possible)_
1. → [draft response or response strategy]
2. → ...

## 🟡 Respond Soon ([N])
| # | From | Subject | By When | Source |
|---|------|---------|---------|--------|
| 1 | [name] | [subject] | [suggested deadline] | [link] |

## ⚡ Quick Reply ([N])
| # | From | Subject | Suggested Reply | Source |
|---|------|---------|-----------------|--------|
| 1 | [name] | [subject] | "Thanks, acknowledged" | [link] |

## 📚 Logged to Compendium ([N])
| Topic | Entries Added | Key Takeaway |
|-------|-------------|--------------|
| [topic] | [N] | [one-line summary of what was learned] |

## ⚪ Archived as Noise ([N])
[count] messages auto-archived. No action needed.

## 📊 Session Stats
- Total processed: [N]
- Response needed: [N] (🔴 [n] + 🟡 [n] + ⚡ [n])
- Knowledge captured: [N] compendium entries
- New topics discovered: [N]
- Attachments processed: [N]
- Noise filtered: [N] ([%] of total)

Phase 7: LEARN & CONNECT

Control Mode: GENERATE → NARRATE | Entropy: E4 → E3

This phase is what makes the skill get smarter over time.

7A: Cross-Reference Discovery (GENERATE)

After each triage session, scan the compendium for connections:

1. Topic clustering: Are any topics converging? (e.g., "Budget" entries and "Hiring" entries are both about the same headcount decision)
2. People overlap: Who appears across multiple topics? (signals who the key connectors/decision-makers are)
3. Timeline patterns: Is a topic accelerating? (more entries in shorter timeframes = something is happening)
4. Decision chains: Do earlier decisions connect to later ones?

7B: Predictive Signals (GENERATE)

Based on compendium history, surface:

## Predictions & Reminders

### Things You'll Probably Need Soon
- [Topic X] has had [N] entries in [timeframe]. You'll likely be asked
  about this in your next [meeting with person].
  → Compendium summary: [link to topic section]

### Connections You Might Not See
- [Topic A] and [Topic B] are converging: [evidence].
  Consider preparing a unified view.

### Stale Topics (No Updates in 14+ Days)
- [Topic] last updated [date] — still relevant? Open items remain:
  [list of open items from compendium]

### People to Follow Up With
- [Person] has [N] open items across [topics]. Last interaction: [date].

7C: Self-Improvement (NARRATE — compress learnings)

After each session, log:

session_learning:
  timestamp: [date]
  messages_processed: [N]
  triage_accuracy: [how many classifications did the user override?]
  missed_signals: [any important messages initially classified as noise?]
  compendium_health:
    topics_growing: [list]
    topics_stale: [list]
    cross_references_added: [N]
  improvement_proposals:
    - "Sender [X] was classified as noise but user flagged as important.
       Add to {{KNOWN_PRIORITIES}}."
    - "Topic [Y] is fragmenting into sub-topics. Propose restructure."
    - "[N] messages about [topic] had no attachment scanning. Switch to
       scan-and-summarize for this sender."

These learnings feed back into the next session:

• Classification rules improve (fewer misclassifications)
• Topic taxonomy evolves (new topics emerge, old ones merge)
• Priority patterns sharpen (the system learns who matters to you)
• Prediction accuracy improves (correlation patterns strengthen)

🔗 Integration: Meeting Prep Assistant

The compendium is a direct input to meeting-prep-assistant.

When meeting-prep-assistant runs Phase 3 (Pull Relevant Notes) and Phase 5 (Assess Real Purpose), it can query the compendium:

meeting-prep-assistant Phase 3:
  "What does the compendium have on [meeting topic]?"
  → Returns: all entries under that topic, with links, decisions, open items

meeting-prep-assistant Phase 5:
  "What's the recent trajectory of [topic]?"
  → Returns: timeline entries, acceleration signals, predictions

meeting-prep-assistant Phase 2:
  "What do I know about [attendee]?"
  → Returns: people index entry, topics they appear in, last interaction, open items

Result: Meeting prep briefs become dramatically richer because they pull from structured knowledge, not just raw email search.

Bidirectional Value

🔄 Operational Modes

Mode 1: Full Triage (Default)

Run all 7 phases. Process entire lookback window. Ideal for catching up on a backlog.

When: Weekly inbox cleanup, Monday morning, after PTO, after ignoring inbox for too long.

Target: 300 unread → triaged in under 1 hour.

Mode 2: Quick Scan

Run Phases 1-3 + Phase 6 only. Skip knowledge extraction and compendium. Produces action list fast.

When: Mid-day check, looking for fires only.

Target: 50 new messages → action list in 5 minutes.

Mode 3: Compendium Catch-Up

Run Phase 4-5 only against already-read messages. Captures knowledge from emails you've handled but didn't log.

When: You've been manually reading emails and want to capture the knowledge into the compendium retroactively.

Target: Scan last 7 days of read messages → compendium updated.

Mode 4: Compendium Query

No triage. Just query the existing compendium.

When: "What do I know about [topic]?" or "Summarize everything from [person] in the last month."

Target: Answer in seconds from structured compendium data.

Mode 5: Learning Review

Run Phase 7 only. Review cross-references, predictions, and improvement proposals from the compendium.

When: Weekly reflection, pre-planning, or when you sense a topic is getting complex and want to see the full picture.

⚠️ Anti-Patterns

INBOX INTELLIGENCE ANTI-PATTERNS:

✗ "Summarize everything"
  Summarizing 300 emails produces a 300-item list. That's not triage.
  → ANCT diagnosis: NARRATE applied to E1 work. Classify first, then
    only narrate what survived triage.

✗ "Read everything to triage it"
  Reading full emails before classifying is O(n) on the wrong step.
  → ANCT diagnosis: Over-investing in Phase 3. Use subject + sender +
    preview for classification. Full read only on Phase 4 knowledge extraction.

✗ "Flat compendium"
  Dumping all summaries into one file with no structure.
  → The compendium must have: topic sections, cross-references, people
    index, timeline, decisions log. Structure IS the value.

✗ "Compendium as archive"
  Storing everything verbatim from emails.
  → The compendium stores extracted knowledge, not email text. Summaries,
    key facts, decisions, and links. Not copy-paste.

✗ "No links back to source"
  Knowledge entries without links to the original message.
  → Every entry must link back. The compendium is an index, not a replacement.

✗ "One-time use"
  Running triage once and never building on it.
  → The value is compound. Each session adds knowledge, improves
    classification, and strengthens predictions. Use it consistently.

CORRECT PATTERNS:

✓ Classify first (DELEGATE), judge second (NARRATE), extract third (GENERATE)
✓ Compendium is structured, indexed, and cross-referenced
✓ Every entry has a source link
✓ Attachments are at minimum linked, ideally scanned
✓ Phase 7 runs every session — the system learns
✓ Manual reads still get logged (INCLUDE_HANDLED = true)
✓ Compendium feeds meeting-prep-assistant and any querying agent

📚 Example Applications

🌐 Platform Notes

/skill inbox-intelligence

Compliance

• PII Risk: High. Email content, sender names, organizational context, attachments. All processing is session-local. The compendium is stored at

  {{COMPENDIUM_PATH}}

  which should be in a private, secure location (private repo, local machine, encrypted drive). Never store the compendium in a public repo.
• Model Recommendation: Sonnet for full triage (Phase 3-4 judgment + extraction). Haiku for Mode 2 quick scan. Opus for Phase 7 deep cross-reference analysis.
• Human Oversight: Triage decisions are proposals. The user reviews the action list before responding. The compendium is additive — the user can edit, correct, or remove entries at any time. Noise classification can be overridden.

Design Credit

This skill's architecture was designed using the adaptive-workflow-architect meta-skill, applying Adaptive Narrative Control Theory (ANCT) to map each phase to its optimal control mode based on entropy level. It integrates with meeting-prep-assistant as a knowledge source for meeting preparation briefs.