git:20260927.31de767 to git:20260928.d21fb8d

5 added, 4 removed. Audit A to A.

---

name: sales-specialist

sheetId: "2.02"

description: "Business card pipeline — OCR extraction, schema normalization, Zynkr platform contact + company creation, company research, and personalized follow-up email drafting. Takes business-card IMAGES specifically — the OCR pass is the point. When the lead signal already arrives as text (a DM thread, an event feedback row, a website enquiry), use sales-outbound instead; for a WHOLE event survey sheet, use sales-client-sourcing."

category: sales-consultant

project: sales-specialist

platform: claude

status: Done

author: Peter Tu

input: "Business card image(s) — single or batch"

process: "OCR via Claude Vision → schema normalization → user review → Zynkr platform write (contact + company + 名片 note) → company research → follow-up email draft → Gmail save"

output: "Contact + company on the Zynkr platform (platform.zynkr.ai/contacts), a 名片 note holding the rest of the card, and a personalized follow-up email saved to Gmail Drafts"

synergy: []

type: agent

skills: ["sales-follow-up-mail"]

house-style: bound

---

# Sales Specialist

```bash

npx skills add https://github.com/peter-tu-zynkr/zynkr-skill-builder --skill sales-specialist

```

Automate the full post-trade-show workflow: extract contact info from business card images, land them on the Zynkr platform, research each company, and draft personalized follow-up emails. Use this skill after a conference, expo, or networking event where you've collected business cards and want them turned into working CRM records plus drafts ready to send before context fades.

**Not this skill when — the signal is already text.** What this skill owns is the OCR pass

from a photographed card. A DM thread, an event feedback row or a website enquiry goes to

`/sales-outbound`; a WHOLE event survey sheet goes to `/sales-client-sourcing`. Everything

downstream (CRM write, research, follow-up draft) is the same either way — only the parser

differs.

---

## Step 0 — Load Config

Read the config file at:

`./sales-specialist-config.md`

Read the contact schema at:

`./references/contact-schema.md`

Store:

- `EMAIL_SIGNATURE` — signature block from config

- `SCHEMA_FIELDS` — the 12-field list: name, title, company, email, phone, mobile, website, address, linkedin, industry, notes, card_date

- `CRM_PROJECT_ID` — Supabase project id from config (the Zynkr platform)

- `CONTACT_SQL` — the contents of `./references/contact-insert.sql` (read it now; you'll fill its placeholders per contact in Step 4)

**If `CRM_PROJECT_ID` is missing or `TBD`:**

```

⚠️ Zynkr platform project id not configured yet.

Set CRM_PROJECT_ID in sales-specialist-config.md before continuing.

```

Stop and ask the user — without it there is nowhere to write the cards.

---

## Step 1 — Card Intake

Display:

```

---------------------------------------------

📇 Business Card Pipeline

---------------------------------------------

Phase: Card Intake

---------------------------------------------

```

Ask the user:

```

Please share your business card image(s).

You can paste one image or multiple at once for batch processing.

```

Detect intake mode:

- **Single card** — 1 image → proceed to Step 2

- **Batch** — multiple images → process all in Step 2, then review all at once in Step 3

Store images as `CARD_IMAGES[]`.

---

## Step 2 — OCR & Extract

Display:

```

---------------------------------------------

Extracting contact info from card(s)...

---------------------------------------------

```

For each card image, use Claude Vision to extract all visible text and map it to the 12-field schema:

**Extraction rules:**

- Read every element on the card: name, title, company, logos, contact details, addresses, URLs, QR code domain (if visible), social handles

- Map each piece of data to the correct schema field

- If a field is not present on the card, set it to `—`

- For `industry`: infer from company name/website if not explicitly stated; use a short label (e.g., "SaaS", "製造業", "零售", "FinTech")

- For `card_date`: default to today's date unless user specifies otherwise

- If text is ambiguous (e.g., could be phone or fax), use field label context clues on the card

Store extracted data as `CONTACTS[]` — one object per card with all 12 fields.

---

## Step 3 — Schema Review

Display:

```

---------------------------------------------

Review Extracted Contact Info

---------------------------------------------

```

For each contact, display a clean table:

```

| Field | Extracted Value |

|------------|-----------------|

| name | [value] |

| title | [value] |

| company | [value] |

| email | [value] |

| phone | [value] |

| mobile | [value] |

| website | [value] |

| address | [value] |

| linkedin | [value] |

| industry | [value] |

| notes | [value] |

| card_date | [value] |

```

For batch mode, display all contacts in a numbered list (`Contact 1 of N`, etc.).

Ask:

```

Does this look correct?

- Type [OK] to proceed

- Type [EDIT field=value] to correct a field (e.g., EDIT email=david@company.com)

- Type [ADD note=...] to append to the notes field

```

Accept corrections until user types `OK`. Update `CONTACTS[]` with confirmed data.

---

## Step 3.5 — Website Lookup (auto-fill if missing)

For each contact where `website` is `—` or blank:

Run a WebSearch query: `"[company name]" official website`

Parse the top result for the company's official domain. Apply these rules:

- Accept only the company's own domain (skip directories like twincn.com, iyp.com.tw, etc.)

- If a clear official site is found, set `website` to the full URL (e.g., `https://example.com`)

- If no clear official site found, leave `website` as `—`

Display the result inline (no separate confirmation step needed):

```

🌐 Website found: [URL] (auto-filled)

```

or:

```

🌐 Website: not found — left blank

```

Update `CONTACTS[]` with the resolved website before proceeding to the platform write.

---

## Step 4 — Zynkr Platform Write

Display:

```

---------------------------------------------

Writing to the Zynkr platform...

---------------------------------------------

```

The **Zynkr platform is the only destination** for a card. Everything lands at

`https://platform.zynkr.ai/contacts`, where the funnel is actually worked — there

is no spreadsheet copy to keep in sync, and no field is dropped on the way:

| Card field | Lands in |

|---|---|

| name | `crm_contacts.last_name` (whole, as printed) |

| title | `crm_contacts.title` |

| email | `crm_contacts.email` (the dedup key) |

| mobile | `crm_contacts.phone` — the number you'd actually use |

| company | `crm_companies.name` (find-or-create) |

| website | `crm_companies.domain` (scheme/`www.` stripped by the SQL) |

| industry | `crm_companies.industry` |

| address | `crm_companies.address` |

| phone (office) | `crm_companies.phone` — belongs to the company, not the person |

| linkedin · notes · card_date | a `名片` note on the contact (`crm_activities`) |

> Earlier versions of this skill also appended a row to a `B2B Customer List`

> Google Sheet, because `crm_contacts` had no home for 7 of the 12 fields. It has

> since grown `domain` / `industry` / `address` / `phone` on the company side, so

> the platform now holds the whole card and the Sheet write has been retired. The

> old Sheet stays readable as the historical record — just don't append to it.

Run this **autonomously** for every confirmed contact (no extra approval gate —

they already signed off in Step 3). For each contact in `CONTACTS[]`:

**1. Clean the values.** Translate the `—` sentinel (and any blank) to an empty

string for every field before using it. The platform should never store a literal `—`.

**2. Use the name as printed.** The contacts page renders a contact as

`last_name + first_name`, so to show the name exactly as it appears on the card,

put the **whole** name in `{{FULL_NAME}}` (it maps to `last_name`) and leave

`first_name` empty — don't split it into 姓/名. If the card prints both a Chinese

and a romanized name (e.g. `王大明 / David Wang`), use the Chinese name if present,

otherwise the romanized one — and store it whole (`王大明`, or `David Wang`).

**3. Keep both numbers.** Pass the mobile as `{{MOBILE}}` and the office line as

`{{PHONE}}` — they go to the contact and the company respectively, so neither is

lost. If the card shows only one number, pass it as `{{MOBILE}}` and leave

`{{PHONE}}` empty; the SQL falls back so the contact always has a number.

**4. Fill and run the SQL.** Take `CONTACT_SQL` (loaded in Step 0 from

`./references/contact-insert.sql`) and substitute the placeholders, **doubling any

- single quote** (`O'Brien` → `O''Brien`). **Fill `<your-company-contact-email>` (the

- CRM owner the contact is filed under) first. If you cannot resolve it, stop and ask for

- it — never run the insert with the blank still in it.** An unfilled owner fails with

- a NOT NULL error (usually on `workspace_id`).

+ single quote** (`O'Brien` → `O''Brien`).

+ > **`<your-company-contact-email>` (the CRM owner the contact is filed under) is a company fact — read it from Atlas first** (`get_knowledge`, key `company.contact-email`, on the `zynkr-atlas` MCP server; the value is the `value:` line) and put the value into the SQL before running it. If Atlas cannot be reached, the key is missing, or there is no `value:` line, ask for it; never run the insert with the blank still in it.

+ An unfilled owner fails with a NOT NULL error (usually on `workspace_id`).

| Placeholder | Value |

|------------------|-------|

| `{{FULL_NAME}}` | the name as printed (Chinese side if bilingual) |

| `{{EMAIL}}` | `email` (empty string if the card had none) |

| `{{TITLE}}` | `title` |

| `{{MOBILE}}` | `mobile` (or the only number on the card) |

| `{{PHONE}}` | `phone` — the office line (empty if the card has one number) |

| `{{COMPANY}}` | `company` (empty string if blank) |

| `{{WEBSITE}}` | `website` as printed — pass the full URL, the SQL normalizes it |

| `{{ADDRESS}}` | `address` |

| `{{INDUSTRY}}` | `industry` (inferred in Step 2 if not printed) |

| `{{LINKEDIN}}` | `linkedin` |

| `{{NOTES}}` | `notes` |

| `{{CARD_DATE}}` | `card_date` (`YYYY-MM-DD`) |

Run it with `mcp__supabase__execute_sql`, `project_id` = `CRM_PROJECT_ID`. One

statement does all of it: find-or-create the company, backfill any company column

that is still empty (**never** overwriting a value a human curated), insert the

contact **only if the email isn't already on the platform** with the baked-in

defaults (`lifecycle_stage=lead`, `legal_basis=consent`, `lead_status=other`,

owner = Peter), and attach the 名片 note.

**5. Read the result and report.**

- One row back (a `contact_id`) → created. Display:

```

✓ [Contact Name] @ [Company] — added → https://platform.zynkr.ai/contacts/[contact_id]

```

- Zero rows back → a contact with this email already exists (de-duped). Display:

```

⏭️ [Contact Name] — already on the platform (matched by email), skipped

```

Note this also means **no note was written** for that card — the note is attached

to a newly created contact only, so a re-scan never stacks duplicate notes.

If the insert errors, surface it, keep the card in `CONTACTS[]` so it can be

retried at the end, and continue with the rest of the pipeline — the follow-up

email shouldn't be blocked by a platform hiccup. Because this is now the only

destination, an unretried failure means the card is **not recorded anywhere**: say

so explicitly in the Step 8 summary rather than reporting the contact as processed.

---

## Step 5 — Company Research

Display:

```

---------------------------------------------

Researching Company...

---------------------------------------------

[Company Name]

```

For each contact, run 3 WebSearch queries in sequence:

1. **Company overview**: `"[company name]" company overview products services`

2. **Recent news**: `"[company name]" news 2024 OR 2025`

3. **Industry context**: `[industry] Taiwan market trends 2025` (or adjust region based on company location)

Synthesize the search results into a `RESEARCH_SUMMARY` for each contact:

```

## Company: [Company Name]

### Overview

[2–3 sentences: what they do, business model, size/stage]

### Recent News

- [key item 1]

- [key item 2]

### Personalization Hook

[The single most relevant/interesting detail to reference in the follow-up email]

```

Display the research summary to the user. Ask:

```

Research looks good? [OK] or [EDIT] to add/change context.

```

---

## Step 6 — Follow-up Email Draft

Display:

```

---------------------------------------------

Drafting Follow-up Email...

---------------------------------------------

[Contact Name] @ [Company]

```

For each contact, launch the `followup-email-writer` agent using the Agent tool.

Pass the following in the prompt:

```

## Contact Info

[all 12 fields from CONTACTS[i]]

## Company Research Summary

[RESEARCH_SUMMARY for this contact]

## Meeting Context

[CONTACTS[i].notes — or "No specific notes" if empty]

## Language

[Detect from card: if Chinese characters on card → zh-TW; otherwise → en]

## Email Signature

[EMAIL_SIGNATURE from config]

```

Wait for the agent to complete. Display the drafted email.

Ask:

```

Email looks good?

- [OK] → save to Gmail Drafts

- [SEND] → send immediately

- [EDIT subject=...] / [EDIT body=...] → adjust the email

- [SKIP] → skip this contact's email

```

Accept edits until user approves.

---

## Step 7 — Save to Gmail Drafts / Send

For each approved email:

**If user chose [OK] (save as draft):**

Use `mcp__gmail__draft_email`:

- `to`: contact email from `CONTACTS[i]`

- `subject`: subject line from agent output

- `body`: email body from agent output

- `from`: `<your-google-workspace-account>`

**If user chose [SEND]:**

Use `mcp__gmail__send_email`:

- `to`: contact email from `CONTACTS[i]`

- `subject`: subject line from agent output

- `body`: email body from agent output

- `from`: `<your-google-workspace-account>`

Display confirmation per contact:

```

✓ [Contact Name] — email saved to Drafts / sent

```

---

## Step 8 — Pipeline Summary

Display final completion summary:

```

=============================================

📇 Business Card Pipeline — Complete

=============================================

Processed: [N] contact(s)

[Contact 1 Name] @ [Company]

✓ Added to the Zynkr platform (or ⏭️ already existed / ✗ FAILED — not recorded)

✓ 名片 note attached

✓ Email drafted / sent

[Contact 2 Name] @ [Company]

✓ Added to the Zynkr platform (or ⏭️ already existed / ✗ FAILED — not recorded)

✓ 名片 note attached

✓ Email saved to Drafts

...

Contacts: https://platform.zynkr.ai/contacts

=============================================

```

If any card shows `✗ FAILED`, list those cards again underneath and offer to retry

the platform write — a failed card exists nowhere but the image it came from.

---

## Progress Indicator

Show this line at the top of each phase:

```

📇 Intake → Extract → Review → Website → Platform → Research → Email → Done

```

Use `▶` for current phase, `✓` for completed, `○` for upcoming.

---

## Error Handling

- If OCR fails to extract a field: mark it `?` and flag it in the review table

- If the platform write fails: surface the error and continue the pipeline, but keep the card flagged. It is the only destination, so a failed card is recorded nowhere — always offer to retry at the end, and never report it as processed.

- If WebSearch returns no results: note "No recent news found" and proceed to email with overview only

- If Gmail draft/send fails: display the error; offer to copy email text to clipboard instead

- Never silently skip a step — always surface errors clearly

## House style

Writing style is **not owned by this file**. The house voice lives in two Google Docs under

`[@] 寫作指南` (`12DBdFz3SK22ie9im_ThFMI7IBRXsTZsV`), read at runtime:

- 《[2.0] Zynkr 通用風格指南 House Voice》 `10bOIQwRm9Pxwgct4hlwCwK_B4Pipai1HqBPZKzyRHSE` —

the universal core, plus the addendum for this surface

- 《[3.2] 禁用詞清單 Forbidden Words》 `1N5sHLP4qzmmhpCGsi6KElxi1z0MFe4QZ0Q_35T10Uyg`

Read both before producing client- or reader-facing text, and scan the draft against 《[3.2]》

before handing it over. If Drive is unreachable, say so in the output rather than proceeding

unchecked. Never re-implement either list inside this file.