Skip to main content

Teaching Email Intelligence

Your inbox shows 47 unread messages. Buried in that count is a 23-message thread about the Q4 budget that you need to respond to before 3 PM. But you can't just read the latest message—critical context is scattered across replies from five different stakeholders over two weeks. Someone made a decision about headcount. Someone assigned you an action item. Someone asked a question that's still unanswered.

You could spend 20 minutes re-reading the entire thread. Or you could invoke a skill that does it in seconds.

This lesson builds an email-summarizer skill that transforms long email threads into actionable intelligence: decisions made, tasks assigned, questions pending. More importantly, you'll learn how skills chain together—how the summarizer's output becomes the drafter's input, creating an automated read-to-respond workflow.

The Information Extraction Challenge​

Long email threads bury critical information in conversational noise. Consider a typical 15-message thread:

Message TypePercentageValue
Greetings and sign-offs25%Zero
Context repetition20%Low (already known)
Discussion and exploration30%Medium (background)
Decisions and action items15%High (actionable)
Open questions10%Critical (blocks progress)

Only 25% of thread content matters for your response. The email-summarizer skill extracts that 25% while discarding the rest.

Why this matters beyond time savings:

  • Accuracy: Manual scanning misses items. Systematic extraction catches everything.
  • Consistency: Same extraction pattern every time, not dependent on your attention span.
  • Handoff: Clear summary enables delegation. Anyone can pick up the thread.
  • Memory: Summaries become searchable records of decisions made.

Understanding Thread Structure​

Before extracting information, you need to understand how email threads are structured. Different email clients format threads differently, but common patterns exist.

Thread Markers to Recognize:

On Mon, Jan 15, 2024 at 2:30 PM Sarah Chen <sarah@company.com> wrote:
From: Marcus Thompson <marcus@company.com>
Sent: Tuesday, January 16, 2024 9:15 AM
To: Team <team@company.com>
Subject: Re: Q4 Budget Discussion
---------- Forwarded message ---------
From: Finance Team <finance@company.com>
Date: Wed, Jan 17, 2024 at 11:00 AM

What to Extract from Structure:

ElementWhat It Tells You
Sender nameWho said what
TimestampWhen (chronological reconstruction)
Subject changesTopic pivots or forwards
Reply depthConversation branches
CC additionsStakeholder expansion

Message Boundary Detection:

Threads nest replies inside replies. The skill needs to identify where one message ends and another begins. Key signals:

  • Timestamp patterns (On [date], [name] wrote:)
  • Quote markers (> at line start)
  • Horizontal rules or separator lines
  • Header blocks (From/To/Subject/Date)

Extraction Targets: What to Pull from Threads​

Your skill needs to extract three categories of information, each requiring different recognition patterns.

1. Decisions Made​

Explicit agreements or conclusions that close discussion:

Signal Patterns:

  • "We've decided to..."
  • "Let's go with..."
  • "Agreed—we'll..."
  • "Final decision:"
  • "Approved by..."
  • "Confirmed:"
  • "Sign-off received"

What to Capture:

  • The decision itself
  • Who made or approved it
  • When it was made
  • Any conditions attached

Example Extraction:

Thread excerpt:
"After reviewing the proposals, let's go with Vendor B for the
cloud migration. Sarah approved the budget yesterday."

Extracted:
DECISION: Select Vendor B for cloud migration
WHO: Team decision, budget approved by Sarah
WHEN: [date of message]
CONDITIONS: None stated

2. Action Items​

Tasks assigned to specific people with ownership:

Signal Patterns:

  • "Can you please..."
  • "[Name] will..."
  • "I'll take care of..."
  • "Action item:"
  • "TODO:"
  • "Next step:"
  • "By [date], we need..."
  • "Your action:"

What to Capture:

  • The task description
  • Who is responsible (owner)
  • Due date (if stated)
  • Dependencies (if mentioned)

Example Extraction:

Thread excerpt:
"Marcus, can you pull the usage metrics from Q3?
We need them before Friday's meeting."

Extracted:
ACTION: Pull Q3 usage metrics
OWNER: Marcus
DUE: Before Friday meeting
DEPENDENCY: Needed for Friday discussion

3. Open Questions​

Unresolved items waiting for response:

Signal Patterns:

  • Lines ending with ?
  • "What do you think about..."
  • "Need input on..."
  • "Waiting for..."
  • "TBD:"
  • "Open question:"
  • "Has anyone..."
  • "Can someone clarify..."

What to Capture:

  • The question itself
  • Who asked it
  • Who should answer (if specified)
  • How long it's been open

Example Extraction:

Thread excerpt (from 3 days ago):
"Does the new pricing apply to existing customers,
or just new signups? Need to know before we update the FAQ."

Extracted:
QUESTION: New pricing scope (existing vs new customers)
ASKED BY: [sender]
ANSWER NEEDED FROM: [not specified - likely product/sales]
OPEN FOR: 3 days
BLOCKS: FAQ update

The Email Summarizer Skill Structure​

Your skill follows the same pattern as email-drafter and email-templates:

.claude/skills/
└── email-summarizer/
    ├── SKILL.md                    # Main skill instructions
    └── references/
        └── extraction-patterns.md  # Pattern recognition library

Building the SKILL.md​

Here's the complete SKILL.md for your email-summarizer skill. Create this at .claude/skills/email-summarizer/SKILL.md:

---
name: email-summarizer
description: This skill summarizes email threads and extracts key information. Use when the user has a long email thread to understand, needs to identify action items, or wants context before replying. Extracts decisions made, action items, and open questions.
---

# Email Summarizer

## Overview

Transform long email threads into actionable summaries with extracted decisions, action items, and context.

## When to Use This Skill

- Email thread has 5+ messages
- Need to quickly understand thread state
- Preparing to reply to a thread
- Extracting action items from discussions
- Creating meeting notes from email exchanges
- Briefing someone on thread status

## Extraction Targets

### 1. Decisions Made

Explicit agreements or conclusions:
- Who decided what
- When it was decided
- Any conditions attached

### 2. Action Items

Tasks assigned to specific people:
- WHO is responsible
- WHAT they need to do
- WHEN it's due (if stated)

### 3. Open Questions

Unresolved items needing response:
- Questions asked but not answered
- Items waiting for input
- Blockers mentioned

### 4. Key Context

Background for response:
- Current state of discussion
- Positions of key stakeholders
- Recent developments

## Output Formats

### Executive Summary (Default)

Quick overview for busy professionals:
- 3-5 sentence summary
- Top 3 action items
- Immediate next step

### Detailed Breakdown

For complex threads:
- Decision log with timestamps
- Full action item list with owners
- Open questions with context
- Stakeholder positions

### Response Context

When preparing to reply:
- What the thread is asking of you
- Your pending action items
- Key points to address

## Workflow

1. Parse thread structure (identify messages and senders)
2. Identify participants and their roles
3. Extract decisions chronologically
4. Extract action items with ownership
5. Identify open questions
6. Format based on user need
7. Offer to chain with /email-drafter for response

## Integration with Other Skills

- Chains with `/email-drafter` for response generation
- Uses tone from `email-drafter/references/tone-guidelines.md`
- Can feed `/email-templates` for structured replies

Key design elements:

  • Clear activation triggers: "5+ messages", "action items", "preparing to reply"
  • Three extraction targets: Decisions, actions, questions (DAQ pattern)
  • Multiple output formats: Different needs require different summaries
  • Skill chaining: Explicit integration with drafter and templates

Creating the Extraction Patterns Reference​

The extraction patterns file gives Claude a recognition library. Create this at .claude/skills/email-summarizer/references/extraction-patterns.md:

# Extraction Patterns for Email Analysis

## Decision Signals

Look for these patterns indicating a decision:
- "We've decided to..."
- "Let's go with..."
- "Agreed - we'll..."
- "Final decision:"
- "Confirmed:"
- "Approved by..."
- "Sign-off received"
- "Moving forward with..."
- "After discussion, we'll..."
- "Consensus is..."

## Action Item Signals

Patterns indicating tasks:
- "Can you please..."
- "I'll take care of..."
- "[Name] will..."
- "Action item:"
- "TODO:"
- "Next step:"
- "By [date], we need..."
- "Your action:"
- "Please send..."
- "Make sure to..."
- "Don't forget to..."

## Question Patterns

Unresolved items:
- Lines ending with "?"
- "What do you think about..."
- "Need input on..."
- "Waiting for..."
- "TBD:"
- "Open question:"
- "Has anyone..."
- "Can someone clarify..."
- "Does anyone know..."
- "Should we..."

## Thread Structure Markers

Identify message boundaries:
- "On [date], [name] wrote:"
- "From: / To: / Subject:"
- "---------- Forwarded message"
- "Begin forwarded message:"
- Timestamp patterns (various formats)
- Quote markers (> at line start)
- Horizontal separator lines

## Participant Roles

Identify key players:
- **Original sender** (initiator) - started the thread
- **Decision makers** (approvers) - can authorize
- **Subject matter experts** (info providers) - provide data
- **Blockers** (waiting on input) - thread stalls on them
- **FYI recipients** (CC'd) - informed but not active

## Urgency Indicators

Signals requiring immediate attention:
- "URGENT:" or "ASAP"
- "EOD" or "End of day"
- "Before [meeting/deadline]"
- "Blocking [something]"
- Exclamation marks (multiple)
- ALL CAPS sections
- "Critical" or "Priority"

## De-Prioritization Signals

Safe to defer:
- "When you get a chance..."
- "No rush, but..."
- "Low priority"
- "FYI only"
- "For your awareness"

Why pattern libraries work:

  • Consistency: Same patterns recognized every time
  • Coverage: Captures variations you might miss
  • Improvement: Add patterns as you encounter new ones
  • Transfer: Patterns work across domains, not just email

Output Format Examples​

Different situations require different summary formats. Here's how each works:

Executive Summary Format​

For quick decisions about whether to engage deeply:

## Thread Summary: Q4 Budget Discussion

**Status**: Active discussion, awaiting final approval

**Key Points** (3):
1. Vendor B selected for cloud migration (approved by Sarah, Jan 15)
2. Headcount frozen at current levels for Q4
3. Marketing budget reduced 15% per CFO guidance

**Your Action Items**:
- Pull Q3 usage metrics before Friday
- Review revised timeline from Marcus

**Immediate Next Step**: Respond to open question about contractor budget

Detailed Breakdown Format​

For complex threads requiring full context:

## Thread Analysis: Q4 Budget Discussion
**Messages**: 23 | **Participants**: 5 | **Duration**: 12 days

### Decisions Log

| Date | Decision | Decided By | Conditions |
|------|----------|------------|------------|
| Jan 15 | Vendor B for migration | Sarah (budget), Team (technical) | Subject to contract review |
| Jan 18 | Headcount freeze | CFO directive | Through Q4 only |
| Jan 22 | Marketing -15% | CFO + CMO | Reallocate to digital |

### Action Items

| Task | Owner | Due | Status |
|------|-------|-----|--------|
| Q3 metrics | Marcus | Friday | Pending |
| Timeline revision | Marcus | Next week | In progress |
| Contract review | Legal | Jan 30 | Not started |
| FAQ update | You | After pricing clarity | Blocked |

### Open Questions

1. **Contractor budget scope** (Jan 20, asked by PM)
   - Does freeze apply to existing contractors?
   - No response yet (3 days open)
   - Blocks: Resource planning

2. **New pricing for existing customers** (Jan 19, asked by Sales)
   - Grandfather existing or apply to all?
   - Blocks: FAQ update, customer comms

### Stakeholder Positions

- **CFO**: Focused on cost reduction, firm on headcount
- **CMO**: Accepted marketing cut, pushing for digital reallocation
- **PM**: Concerned about timeline with frozen resources
- **Sales**: Waiting on pricing clarity for customer communication

Response Context Format​

When you're about to reply:

## Response Context: Q4 Budget Thread

**What the thread is asking of you**:
1. Pull Q3 usage metrics (explicit request from Marcus)
2. Clarify contractor budget question (you're likely the right person)
3. Confirm timeline feasibility with reduced resources

**Your pending action items**:
- Q3 metrics: Due Friday (2 days)
- FAQ update: Blocked on pricing question

**Key points to address in your reply**:
- Confirm you'll deliver metrics by Friday
- Provide contractor budget guidance (or defer to CFO)
- Flag timeline risk if resources stay frozen

**Tone note**: CFO is in thread—be concise, data-focused

Skill Chaining: Summarizer to Drafter​

The real power emerges when skills work together. Your email-summarizer produces context that your email-drafter consumes:

The Chain:

[Long Thread] → /email-summarizer → [Context] → /email-drafter → [Reply Draft]

How It Works:

  1. Invoke summarizer on the thread
  2. Summarizer extracts action items, decisions, questions
  3. Ask to draft a response
  4. Drafter uses summary as context
  5. Draft addresses extracted items with your tone

Example Flow:

User: Summarize this thread and then help me respond.

[Claude invokes /email-summarizer]

Summary:
- Decision: Vendor B selected
- Action for you: Provide Q3 metrics by Friday
- Open question: Contractor budget (needs your input)

[Claude recognizes response intent, invokes /email-drafter]

Using your summary, here's a draft response:

Hi team,

Quick updates on the open items:

**Q3 Metrics**: I'll have these ready by Thursday EOD,
giving us a day buffer before Friday's meeting.

**Contractor Budget**: For existing contractors, I recommend
we honor current agreements through Q4. New contractor
requests should route through the approval exception process.

Let me know if you need anything else before Friday.

Best,
[Your name]

Why chaining matters:

  • Context preservation: Summary context flows into draft
  • Efficiency: Single workflow, not separate prompts
  • Consistency: Same tone guidelines apply
  • Completeness: Draft addresses all extracted items

Testing Your Skill​

Let's verify the skill works correctly.

Step 1: Create the directory structure

mkdir -p .claude/skills/email-summarizer/references

Output:

(no output - directories created silently)

Step 2: Verify your structure

find .claude/skills -type d | sort

Output:

.claude/skills
.claude/skills/email-drafter
.claude/skills/email-drafter/references
.claude/skills/email-summarizer
.claude/skills/email-summarizer/references
.claude/skills/email-templates
.claude/skills/email-templates/templates

Step 3: Create the files

Copy the SKILL.md content to .claude/skills/email-summarizer/SKILL.md and the extraction patterns to .claude/skills/email-summarizer/references/extraction-patterns.md.

Step 4: Test with a sample thread

Provide Claude with a multi-message email thread and invoke the skill:

/email-summarizer

[Paste your email thread here]

Give me an executive summary and list my action items.

Expected Behavior:

  1. Claude recognizes thread parsing is needed
  2. Identifies message boundaries and senders
  3. Extracts decisions, actions, questions
  4. Formats as executive summary
  5. Highlights your specific action items

Common Extraction Mistakes​

When building extraction logic, these patterns cause problems:

Mistake 1: Missing implied decisions

Thread: "I don't hear any objections, so we'll proceed with Plan A."

WRONG: No decision extracted (no explicit "decided" word)
RIGHT: Decision extracted (consensus by silence)

Mistake 2: Confusing suggestions with action items

Thread: "We could also consider reaching out to Partner X."

WRONG: Action item - reach out to Partner X
RIGHT: Suggestion only (no ownership, no commitment)

Mistake 3: Missing conditional action items

Thread: "If the budget clears, Sarah will start procurement."

WRONG: Action item - Sarah starts procurement
RIGHT: Conditional action - Sarah starts IF budget clears

Mistake 4: Ignoring thread chronology

Early message: "Let's use Vendor A"
Later message: "After review, switching to Vendor B"

WRONG: Decision - use Vendor A
RIGHT: Decision - use Vendor B (supersedes earlier)

Extending the Skill​

Once the basic skill works, consider these enhancements:

ExtensionValueComplexity
Sentiment detectionIdentify tension or urgencyMedium
Stakeholder mappingAuto-identify rolesLow
Timeline extractionBuild project timeline from threadMedium
Commitment trackingMatch promises to follow-throughHigh

Sentiment example addition to SKILL.md:

## Sentiment Indicators

### Urgency Signals
- Multiple exclamation marks
- "URGENT", "ASAP", "Critical"
- Short response times (< 1 hour between messages)

### Tension Signals
- Defensive language ("As I mentioned...")
- Escalation (new executives added to CC)
- Formal tone shift (first names → full names)

Try With AI​

Setup: Open Claude Code in the email-assistant directory with the email-summarizer skill installed.

Prompt 1: Basic Thread Extraction

/email-summarizer

Here's a thread I need to understand quickly:

[Paste a real or sample email thread with 5+ messages
that contains at least one decision, one action item,
and one open question]

Give me the executive summary format.

What you're learning: This tests the core extraction workflow. Review the output: Did Claude correctly identify message boundaries? Are the extraction targets (decisions, actions, questions) accurate? Did anything get missed or miscategorized?

Prompt 2: Response Context Generation

Now give me the response context format for that same thread.
I need to understand what the thread is asking of ME specifically
and what I should address in my reply.

What you're learning: This tests output format flexibility. The same thread should produce different summaries for different purposes. Compare the executive summary to the response context—same source data, different focus.

Prompt 3: Chain to Email Drafter

Based on that summary, help me draft a response.
Use /email-drafter to compose a reply that addresses
all the action items assigned to me and answers
the open questions I can answer.

What you're learning: This demonstrates skill chaining. The summarizer's output becomes the drafter's context. Notice how the draft addresses the specific items extracted from the summary. This is workflow automation—multiple skills coordinating automatically.

Safety Note: When summarizing threads containing sensitive information, be mindful of what you share with AI systems. Summarization is powerful but should be used appropriately for the content's sensitivity level.