linkedin-outreach（链接外展）

LinkedIn Outreach

The LinkedIn counterpart to cold-email-outreach. Takes qualified leads from Supabase, builds personalized LinkedIn message sequences, exports for the user's LinkedIn outreach tool, and logs everything back to Supabase.

Tool-agnostic: Asks the user which LinkedIn tool they use. All tools are CSV-import based — no API/MCP automation for LinkedIn tools (they're browser-based). Adapters handle column mapping and format differences per tool.

When to Auto-Load

Load this skill when:

• User says "LinkedIn outreach", "connect with these leads on LinkedIn", "send LinkedIn messages", "set up a LinkedIn campaign"
• An upstream skill connects with "create LinkedIn campaign" or "passes: supabase-eligible-leads" and user specifies LinkedIn
• User completes lead-qualification and wants to reach out via LinkedIn

Supported Outreach Tools

This skill does NOT assume a specific tool. It asks first, then adapts.

Tool selection logic:

1. Ask user in Phase 0: "Which LinkedIn outreach tool do you use?"
2. Generate tool-specific import CSV based on selection
3. If Other or unknown → generate generic CSV (linkedin_url, first_name, last_name, company, title, connection_request, followup_1, followup_2, followup_3, inmail_subject, inmail_body) and ask user for their tool's import requirements

Prerequisites

Supabase

People must be stored in Supabase with the schema from tools/supabase/schema.sql. The people and outreach_log tables must exist. Run python3 tools/supabase/setup_database.py if setting up fresh.

Environment variables in .env:

SUPABASE_URL=https://xxx.supabase.co
SUPABASE_SERVICE_ROLE_KEY=eyJ...

LinkedIn Tool

Just need CSV export — no API keys required. The user imports the CSV into their tool manually.

Character Limits

LinkedIn enforces strict character limits. All generated messages must respect these.

Enforcement: After generating any message, count characters. If over the limit, rewrite — do not truncate. Truncated messages look broken.

Phase 0: Intake

Ask all questions at once. Organize by category. Skip any already answered by an upstream skill.

Campaign Goal

1. What's the objective? (book meetings, drive demo requests, get replies, build relationships, nurture)
2. What's the outreach angle or hook? (hiring signal, competitor displacement, event-based, pain-based, cold database, KOL engagement, mutual connection)
3. What should we name this campaign?

Outreach Tool

1. Which LinkedIn outreach tool do you use? (Dripify / Botdog / Expandi / PhantomBuster / Other / Just give me a CSV)

Lead Selection

1. Which leads should we target? Options:

- All leads for a specific client_name - Specific icp_segment - Title patterns (e.g., "VP Operations", "Director of Sales") - Industry or location filters - qualification_score above a threshold - Specific source (crustdata, apollo, linkedin, etc.) - Custom filter (describe what you want)

1. Any exclusions? (specific companies, recently contacted leads, certain titles)
2. Max campaign size? (default: 100 — LinkedIn tools have lower daily limits than email)

Tone & Style

1. Which tone preset? Present these options:

- Casual Professional — Friendly, human, slightly informal. Like messaging a peer. (default) - Thought Leader — Lead with insight or a contrarian take. Position sender as an expert. - Provocative — Challenge assumptions, pattern-interrupt. Higher risk, higher reward. - Enterprise Formal — Polished, structured. For regulated industries or C-suite targets. - Custom — Paste reference messages that worked before, or describe the vibe.

1. Any reference messages that have worked well? (paste examples — these override tone presets)

Sequence Structure

1. How many follow-ups after connection? (default: 3)
2. Timing between messages? (default: Day 0 connection / Day 3 FU1 / Day 7 FU2 / Day 14 FU3)
3. Include InMail as a separate step for leads who don't accept the connection? (default: yes)

Personalization

1. What signal data is available for these leads? (comment text, post they engaged with, mutual connections, hiring signals, event attendance)
2. Any proof points or case studies to reference? (customer names, metrics, testimonials)

Phase 1: Lead Selection from Supabase

Connect

Use the shared Supabase client:

import sys, os
sys.path.insert(0, os.path.join("tools", "supabase"))
from supabase_client import SupabaseClient

client = SupabaseClient(os.environ["SUPABASE_URL"], os.environ["SUPABASE_SERVICE_ROLE_KEY"])

Build Filters

Map user criteria to PostgREST query parameters on the people table:

Critical: For LinkedIn outreach, people MUST have a linkedin_url. Filter out people without one — they can't be contacted via LinkedIn.

Cooldown Filter (Mandatory)

Always exclude people contacted within 84 days (12 weeks) on ANY channel (email or LinkedIn). This is not optional.

Use the shared client's check_cooldown() method:

in_cooldown = client.check_cooldown(client_name="happy-robot", cooldown_days=84)
# Returns set of person_id strings still in cooldown

Or query directly:

1. Query outreach_log for person_ids with sent_date in the last 84 days: GET /rest/v1/outreach_log?select=person_id&sent_date=gte.{84_days_ago}&status=neq.bounced&client_name=eq.{client}
2. Collect those person_ids into an exclusion set
3. Add id=not.in.({excluded_ids}) to the people query

Note: Cooldown applies across channels. A person emailed 30 days ago is still in cooldown for LinkedIn. This prevents multi-channel bombardment.

Present & Confirm

Show a sample table (10-15 leads) with:

• Name, Title, Company, Industry, Score, LinkedIn URL, Last Contacted, Signal Type

Tell user: total eligible leads, how many excluded by cooldown, how many excluded for missing LinkedIn URL.

Ask user to confirm or adjust filters before proceeding.

Phase 2: Sequence Design

Present the sequence plan as a table before writing any copy:

*InMail is sent to leads who haven't accepted the connection request by Day 7.

Key differences from email sequences:

• Connection request is the gatekeeper — it must earn the accept. No selling in the connection note.
• Follow-ups are conversational, not broadcast. They should read like DMs, not emails.
• No subject lines except for InMail.
• Shorter is almost always better. A 2-sentence message outperforms a 5-sentence one on LinkedIn.

Get user approval on the structure before generating copy in Phase 3.

Phase 3: Message Generation

Generate messages directly in this skill (no external sub-skill needed — LinkedIn messages are short enough to handle inline).

Signal-Aware Template Selection

Select the appropriate sequence template based on lead signal data:

Tone Calibration

1. Load the selected tone preset from templates/tone-presets.json
2. If user provided reference messages, those override the preset — analyze the reference messages for tone, length, structure, and vocabulary
3. Apply tone guidelines to all generated messages

Calibration Loop

1. Generate sample messages for 3-5 leads first (pick leads with different signal richness levels)
2. Present to user for review — show the full sequence for each sample lead
3. Iterate until approved (max 3 rounds)
4. Batch generate remaining messages after approval

Writing Guidelines

Connection Request (300 chars max):

• Lead with the signal (what they did/said that caught your attention)
• One sentence of relevance (why you're connecting)
• No pitch, no CTA, no "I'd love to..." — just be interesting enough to accept
• MUST be under 300 characters. Count every character.

Follow-up 1 (value-first):

• Thank for connecting (briefly — one clause, not a whole sentence)
• Share something genuinely useful: insight, article, observation about their company/industry
• End with a question, not a pitch

Follow-up 2 (social proof):

• Reference a relevant customer or case study
• Connect it to their specific situation
• Make a specific, low-commitment ask (15-min call, async question)

Follow-up 3 (breakup):

• Acknowledge you've been reaching out
• One-line value recap
• Leave the door open without pressure
• Shortest message in the sequence

InMail (standalone pitch):

• Subject: 200 chars max — curiosity-driven, not salesy
• Body: 1,900 chars max — must work standalone since they haven't accepted your connection
• Include context for why you're reaching out (the signal)
• Must work even if they've never heard of you

Merge Variables

Standard variables available for all leads:

• {first_name}, {last_name}, {company}, {title}, {industry}, {location}

Signal-specific variables (available based on source):

• {comment_snippet} — the text of their LinkedIn comment
• {pain_topic} — the pain point they engaged with
• {competitor_name} — the competitor whose post they engaged with
• {kol_name} — the KOL whose post they engaged with
• {post_topic} — what the post was about
• {event_name} — the event they attended
• {role_hiring_for} — the role they're hiring for
• {job_posting_detail} — a detail from the job posting

Phase 4: Campaign Export

Step 1: Generate Universal CSV

Core columns for all exports:

linkedin_url, first_name, last_name, company, title, connection_request, followup_1, followup_2, followup_3, inmail_subject, inmail_body

Step 2: Format for Selected Tool

Dripify:

• Column mapping: Profile URL → linkedin_url, Note → connection_request, Message 1 → followup_1, etc.
• Dripify expects one row per lead with all messages in separate columns
• Export format: CSV with headers matching Dripify's import template

Botdog:

• Column mapping: linkedin_profile_url → linkedin_url, connection_note → connection_request, message_1 → followup_1, etc.
• Export format: CSV

Expandi:

• Column mapping: LinkedIn URL → linkedin_url, Connection message → connection_request, Follow-up #1 → followup_1, etc.
• Supports InMail columns: InMail subject, InMail message
• Export format: CSV

PhantomBuster:

• Column mapping: profileUrl → linkedin_url, message → connection_request
• PhantomBuster typically handles one action at a time — may need separate CSVs for connection + follow-ups
• Export format: CSV

Manual / Other:

• Use the universal CSV format
• Provide column descriptions and tool-agnostic import instructions
• Ask user what format their tool expects, adjust if needed

Step 3: Save Files

skills/linkedin-outreach/output/{campaign-name}-{YYYY-MM-DD}.csv

Create the output/ directory if it doesn't exist.

Step 4: Optional Google Sheet

If user wants a review sheet, use google-sheets-write capability to create a sheet with:

• Tab 1: Lead list with all messages (one row per lead)
• Tab 2: Sequence templates (the master templates used)
• Tab 3: Campaign config summary

Phase 5: Review & Approval

Present campaign summary:

Campaign: {name}
Tool: {dripify/botdog/expandi/etc.}
Leads: {count}
Sequence: Connection + {followup_count} follow-ups + InMail
Timing: Day 0 → Day {last_day}
Tone: {preset_name}
Signal types: {breakdown by signal type}
Leads with rich signal: {count} ({percentage}%)
Leads with lean signal: {count} ({percentage}%)
Export file: {file_path}

Hard Approval Gate

Do NOT mark the campaign as ready without explicit user confirmation. Present the summary, then ask: "Ready to finalize? Type 'yes' to mark as ready for import."

After approval:

• Tell user the file is ready for import into their LinkedIn tool
• Provide the file path
• Give tool-specific import instructions (see Phase 4)
• Remind user to verify the first 5-10 messages look correct after import

Phase 6: Logging

Database Write Policy

All database writes in this phase require the user's prior approval from the finalization gate in Phase 5. Since LinkedIn campaigns are always exported (never auto-launched), confirm with the user before logging to outreach_log — they may not have actually imported the campaign into their LinkedIn tool yet. Only log after the user confirms the export is final.

Log to Supabase

After export and user confirmation, insert records into outreach_log:

POST /rest/v1/outreach_log
Prefer: return=minimal

[
  {
    "person_id": "{person_uuid}",
    "campaign_name": "{campaign_name}",
    "channel": "linkedin",
    "tool": "{dripify/botdog/expandi/phantombuster/manual}",
    "sent_date": "{ISO timestamp}",
    "status": "exported",
    "client_name": "{client_name}"
  },
  ...
]

Or use the shared client:

client.log_outreach(entries)

Status is "exported", not "sent". LinkedIn tools are browser-based — we can't confirm delivery. The status changes to "sent" when the user confirms they launched the campaign in their tool.

Update People Records

Update last_contacted on the people table for all people in this campaign:

PATCH /rest/v1/people?id=in.({person_ids})
{ "last_contacted": "{ISO timestamp}" }

Present Summary

Campaign: {name}
{count} people logged to outreach_log (channel: linkedin)
last_contacted updated for {count} people
Cooldown active until: {date + 84 days}
Next eligible re-contact: {date}
File ready: {file_path}

Cooldown Enforcement Rules

Reference section for cooldown logic used throughout this skill. Shared with cold-email-outreach.

Output Directory

Campaign exports are saved to:

skills/linkedin-outreach/output/

Create this directory if it doesn't exist. Files are named {campaign-name}-{YYYY-MM-DD}.csv.