hermes-skills/email/email-automation/references/continuous-invoice-organizer/learning-json-format.md

learning.json Format

The learning file at ~/.hermes/email-organizer/learning.json accumulates email classification rules across cron runs.

Structure

{
  "corrections": [],
  "patterns": {
    "rechnung": {
      "<pattern_name>": {
        "sender": "exact@sender.com",
        "sender_pattern": "@appleid.com",
        "subject_keywords": ["keyword1", "keyword2"],
        "content_keywords": ["rechnung", "rechnungsnummer"],
        "folder": "Rechnungen_YYYY_MM",
        "source": "Vendor name",
        "notes": "Description and context"
      }
    },
    "bestellbestaetigung": {
      "<pattern_name>": { ... }
    },
    "ignorieren": {
      "<pattern_name>": { ... }
    }
  },
  "structured_ignore_patterns": [],
  "last_scan": "2026-06-10 09:13",
  "total_processed": 554,
  "total_moved": 20
}

Pattern Fields

Field	Required	Description
sender	Conditional	Exact email address (use with sender_pattern only)
sender_pattern	Conditional	Email domain pattern with @ prefix (e.g., @appleid.com)
subject_keywords	Yes	Keywords matched case-insensitively in subject
content_keywords	Conditional	Keywords matched in email body
folder	Conditional	Override target folder (defaults based on category)
source	Yes	Human-readable vendor/source name
notes	Yes	Context: what type of email, whether it's an invoice or order, any caveats

Decision Criteria

Invoice (Rechnung)

Contains a factored charge: Rechnungsnummer, Gesamtbetrag, Steuern, Fälligkeitsdatum, Lastschrift.

Order Confirmation (Bestellbestätigung)

Contains an order number and items but NO invoice details. Invoice arrives separately.

Ignore

• Delivery notifications ("zugestellt", "versendet")
• Newsletters, marketing emails
• Social media notifications
• Job alerts
• Technical reports (solar, DMARC)

Update Workflow

When a new pattern emerges during a scan:

1. Add entry under the correct category in patterns
2. Include source with vendor name
3. Include notes with distinguishing characteristics
4. Increment total_processed and total_moved
5. Update last_scan timestamp