# Email Content Compose

All-in-one skill: fetch leads, compose personalized outreach emails, export as EML and upload to R2.

## IMPORTANT — Execution Rules

- This skill uses **Bun + TypeScript**. Do NOT create Python scripts.
- Do NOT overwrite existing files in this skill directory.
- Use the existing CLI scripts below. Do NOT write your own upload/compose logic.

## Phases

### 1. Fetch (scripts/fetch.ts)
Retrieves the lead-dataset for a completed cold-outreach workflow.

```bash
cd ~/clawd/skills/email-content-compose
bun run fetch -- --workflow-id=<id>
```

Returns: JSON with businesses (name, website, reviews, emails) and summary stats.

### 2. Compose (LLM)
The LLM agent reads fetch output and composes a personalized email per business.
Save drafts as a JSON array file. Each draft must have:
- `recipient_email` (string)
- `recipient_name` (string | null)
- `subject` (string, under 60 chars)
- `body_html` (string, professional HTML with inline styles)
- `body_text` (string, plain text fallback)
- `personalization_context` (object, see below)

#### a) Sender Profile
Before composing, read `sender-profile.json` in this skill directory. Use it to populate sender identity, product info, and signature in every email.

#### b) Language Selection
Auto-select email language based on the `country` field from fetch output:
- us/gb/au/ca → English
- cn → 中文
- jp → 日本語
- kr → 한국어
- de → Deutsch
- fr → Français
- es/mx → Español
- pt/br → Português
- Unlisted countries → default to English

Write the **entire** email (subject, body, signature) in the selected language.

#### c) Review Pain-Point Analysis
When `reviews_data` is non-empty for a business:
1. Parse JSON into a review array
2. Filter for negative reviews / complaints related to the sender's `product_category`
3. Extract 1–2 pain points that the sender's products can solve
4. Analyze at most 10 reviews, prioritizing low-score ones
5. If `reviews_data` is empty or null, fall back to the business's general info (category, rating) to craft the email

#### d) Email Template Structure
1. **Subject** — Under 60 chars, in the target language, referencing a specific pain point or business need
2. **Opening** — Address the business by name; demonstrate familiarity with their operations
3. **Pain Point Bridge** — Reference pain-point patterns from reviews (do NOT quote reviews verbatim); connect them to problems the sender's product solves
4. **Value Proposition** — Introduce sender company and products using `product_highlights` from sender-profile.json
5. **CTA** — Low-friction call-to-action: free samples, catalog, or a brief call
6. **Signature** — Use contact info from sender-profile.json (name, title, email, phone, website)

#### e) Personalization Tracking
Populate the `personalization_context` field on each draft:
```json
{
  "language": "selected language",
  "pain_points_used": ["pain point 1", "pain point 2"],
  "reviews_analyzed": true,
  "sender_product_match": "brief note on how sender product connects to this business"
}
```

#### General Rules
- Tone: professional, consultative, not salesy
- Save the drafts array to a temp JSON file, e.g. `/tmp/drafts-<workflow-id>.json`

### 3. Export (scripts/export.ts)
Converts drafts to RFC 5322 EML files, uploads individual EMLs + ZIP bundle to R2.

```bash
cd ~/clawd/skills/email-content-compose
bun run export -- --drafts=/tmp/drafts-<workflow-id>.json --workflow-id=<id> [--from=<email>] [--dry-run]
```

Returns: JSON with per-file R2 URLs and a bundle.zip URL.

## Full Pipeline Example

```bash
cd ~/clawd/skills/email-content-compose

# 1. Fetch leads
bun run fetch -- --workflow-id=outreach-xxx > /tmp/leads.json

# 2. LLM composes drafts → saves to /tmp/drafts-outreach-xxx.json

# 3. Export EML + upload to R2
bun run export -- --drafts=/tmp/drafts-outreach-xxx.json --workflow-id=outreach-xxx
```

## Config

Auth handled automatically by auth-runtime via `~/.openclaw/.env`.

R2 and email config in `~/.openclaw/.env`:
- `CLOUDFLARE_*`: R2 upload credentials
- `SENDER_EMAIL`: From header in EML files