Gmail Auto-Label

Goal

Fetch inbox emails, classify them via parallel subagents into Action Required / Waiting On / Reference, and apply labels in bulk via Gmail API.

Scripts

• ./scripts/gmail_label_fetch.py - Fetch email summaries as compact JSON
• ./scripts/gmail_label_split.py - Split emails into N chunks for parallel classification
• ./scripts/gmail_label_merge.py - Merge classified chunks into single labels.json
• ./scripts/gmail_label_apply.py - Apply label classifications in bulk

Subagent

• email-classifier — defined in .claude/agents/email-classifier.md
• Model: Sonnet 4.5 (fast, cost-efficient classification)
• Each subagent reads one chunk, writes one classified output file

Flow (Parallel — default)

Step 1: Fetch emails

python3 .claude/skills/gmail-label/scripts/gmail_label_fetch.py \
  --account ACCOUNT --query "in:inbox" --limit 100 --output .tmp/emails.json

Step 2: Split into chunks

python3 .claude/skills/gmail-label/scripts/gmail_label_split.py \
  --input .tmp/emails.json --chunks 10 --output-dir .tmp/chunks

Step 3: Classify in parallel (spawn 10 subagents)

Spawn 10 email-classifier subagents in background, one per chunk. Each subagent:

• Reads .tmp/chunks/chunk_N.json
• Classifies each email
• Writes .tmp/chunks/classified_N.json

Use the Task tool with run_in_background: true and model: "sonnet". Launch ALL 10 in a single message for true parallelism:

For each chunk 0-9, spawn a Task with:
  subagent_type: "email-classifier"
  model: "sonnet"
  run_in_background: true
  prompt: "Read /absolute/path/.tmp/chunks/chunk_N.json, classify each email, write results to /absolute/path/.tmp/chunks/classified_N.json"

CRITICAL: Do NOT use TaskOutput to read subagent results. The subagents write their results to files — the main agent never needs to see the classification data. Reading TaskOutput will flood the context window and cause "prompt too long" errors with large batches (500+ emails).

Instead, poll for file existence:

# Wait until all classified files exist (timeout after 120s)
for i in $(seq 0 9); do
  while [ ! -f ".tmp/chunks/classified_$i.json" ]; do sleep 2; done
done

Then proceed directly to Step 4 (merge).

Step 4: Merge classifications

python3 .claude/skills/gmail-label/scripts/gmail_label_merge.py \
  --input-dir .tmp/chunks --output .tmp/labels.json

Step 5: Apply labels

python3 .claude/skills/gmail-label/scripts/gmail_label_apply.py \
  --account ACCOUNT --input .tmp/labels.json

Classification Guidelines

Action Required:

• Security alerts that need verification
• Expiring credit cards / domain renewals with deadlines
• Slack @mentions asking questions
• New team members to greet (Slack join notifications)
• Client emails needing response
• Business listing updates (Google Business Profile, Bing Places)
• Stripe action-required notices

Waiting On:

• Outbound sales emails awaiting reply
• Support tickets awaiting resolution
• Proposals sent, pending response

Reference:

• Marketing newsletters (DigitalMarketer, etc.)
• Charity/nonprofit newsletters (RAPS, etc.)
• Google Business Profile performance reports
• Promotional offers (Blinkist, sales, etc.)
• Platform update notifications (Google Play, Apify, etc.)
• Confirmation codes (already used)
• Real estate newsletters (Westbank, etc.)
• Gaming account emails (Riot Games, etc.)
• Informational security alerts (2FA turned on, etc.)
• Health advisories
• Legal/policy update notices

Account Registry

Accounts are stored in gmail_accounts.json at workspace root. Each account needs:

• email - Gmail address
• token_file - Path to OAuth token

Adding New Accounts

python3 .claude/skills/gmail-inbox/scripts/gmail_multi_auth.py --account ACCOUNT_NAME --email EMAIL

Performance

• Serial flow: ~36s for 100 emails (fetch 1s + classify 34s + apply 1s)
• Parallel flow: ~30s for 100 emails (classify 19s + merge <1s + apply ~10s)
• Classification is the bottleneck; 10 parallel subagents cut it from 34s to 19s (~1.8x speedup)
• Agent startup overhead (~4s stagger) limits theoretical gains
• Apply step may vary due to Gmail API throttling

---

Schema

Inputs

Name	Type	Required	Description
account	string	Yes	Gmail account name from registry
query	string	No	Gmail search query (default: 'in:inbox')
limit	integer	No	Max emails to classify (default: 100)
chunks	integer	No	Parallel classification chunks (default: 10)

Outputs

Name	Type	Description
labels_applied	object	Count per category: Action Required, Waiting On, Reference

Credentials

Name	Source
credentials.json	file
token_*.json	file (auto-generated)

Composable With

Skills that chain well with this one: gmail-inbox

Cost

Claude Sonnet API for classification