outreach-conductor

A Claude Code skill that runs a weekly outbound cycle like a careful researcher would, just faster. It finds companies that match your ideal customer profile, verifies each company and the right person, researches them, removes anyone you already contacted or met, writes personalized emails in your own voice, checks every one of them, sends them from your own mailbox in small daily batches, picks a short list of people to connect with on LinkedIn, logs everything to your CRM, and learns from last week's reply rates.

It asks its setup questions once. After that it runs the whole week by itself and reports what it did. The only thing it asks about is money: paid data above your budget, or any purchase. If you would rather approve each batch, set autonomy.mode: review. If you have not connected a mail sender, it creates drafts and you send.

Why it exists

Most outbound automation optimizes for volume and gets ignored. This skill is built around the opposite: a verification ladder, a per-person research bar, a dedupe gate across every place you might have talked to someone, drafting rules copied from emails you actually sent, a no-repeat check, and a review loop. Volume is a config setting (default 100 a week). Quality is not.

Requirements

• Claude Code with web search.
• Python 3.9 or newer (standard library only; PyYAML is used if installed).
• Optional: an email connector that can send from your mailbox (or just create drafts), a company database or enrichment tool, a search API, a CRM with a CLI or API. None are required. A CSV file works as the CRM.

Install

git clone https://github.com/mariagorskikh/outreach-conductor ~/.claude/skills/outreach-conductor

Configure

Either let the skill interview you on first run ("run weekly outreach" with no config starts the interview), or do it by hand:

cd ~/.claude/skills/outreach-conductor
cp config.example.yaml config.yaml
mkdir -p data voice
cp examples/sent-examples.example.md voice/sent-examples.md   # then paste your real emails

Then edit config.yaml: your name and voice notes, your offer, your ICP and fit rubric, buckets and angles with their shares, the weekly volume, your CRM, the sender (email.send_via, email.daily_send_cap), the compliance.footer, the LinkedIn cap, the spending budget, and autonomy.mode. Keep API keys in environment variables, never in the config.

Check it parses:

python3 scripts/week.py init

Environment variables:

Variable	What it does
OUTREACH_CONDUCTOR_CONFIG	path to a config file outside the skill folder
OUTREACH_STATE_DIR	where week folders live (default state/ next to the config)

Example prompts

• "Run this week's outreach."
• "Set up outreach-conductor for my company." (starts the interview)
• "Find and draft 50 prospects in the events and warm buckets only."
• "Resume the outreach run." (picks up from the state folder)
• "Who should I connect with on LinkedIn this week?"
• "Switch outreach-conductor to review mode." (approve each batch yourself)
• "What got replies last month? Update next week's plan."

How it works

flowchart LR
  A[Setup questions, once] --> B[Find companies]
  B --> C[Verify and dedupe]
  C --> D[Research and write]
  D --> E[Self-review gate]
  E --> F{Paid data above budget?}
  F -- yes --> G[Ask you first]
  F -- no --> H[Send daily batches]
  G --> H
  H --> I[CRM, LinkedIn list, report]

1. Setup, once. On the first run it asks a short batch of questions (your offer, ICP, voice, sender, caps, budget) and writes config.yaml. Weekly runs ask nothing.
2. Kickoff. Reads the config and your example emails, snapshots the CRM, logs last week's sends, replies and bounces, and reads the reply rates.
3. Find. Parallel scout agents search per bucket and angle (look-alikes of your best customers first, then fresh signals like funding and launches) and write candidates to files as they go.
4. Verify. Each company must be real, funded or staffed, with a working site that sells into your ICP. The right person must have a real profile. Every candidate is checked against your CRM, sent mail, drafts, message threads, calendar and opt-out lists.
5. Enrich. Finds addresses on a ladder from proven to guessed, and keeps at least 80% on a verified rung. Paid credits stay inside your weekly budget; anything above it waits for your yes.
6. Write and self-review. Writers research each person and write in your voice. A no-repeat check, a lint, and a fresh reviewer per slice of 50 catch copy-paste, false hooks and AI tells. Every fact and link is checked before anything goes out.
7. Send. Each business day it sends a batch under your daily cap from your own mailbox, spread through business hours, with your footer, skipping opt-outs and unverified addresses. Real sends are logged to your CRM as touches. No sender configured: drafts.
8. LinkedIn. A ranked list of people to connect with. If you run an invite tool, it gets the list inside your caps; otherwise you send by hand.
9. Report and learn. One summary of everything that went out, what was skipped and why. reply_rates.py measures reply and meeting rates by opener, subject, length, bucket and angle, and next week's briefs change accordingly.

Everything for a week lives in state/<YYYY-Www>/, so a run can stop and resume across sessions without redoing work, and never sends the same email twice.

It does by itself	It asks you first	It always stops on
research, writing, review, sending inside the daily cap, CRM updates, the LinkedIn list, small quota shifts from reply data	paid data above spending.weekly_budget, buying a list, tool, domain or mailbox, upgrading a plan	bounces above email.bounce_stop, platform or mail provider warnings, login or captcha prompts, opt-outs, a fact it cannot verify

여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗