Outbound Agent Skills
29 Claude Skills and 3 scripts for an AI agent that runs B2B outbound on LinkedIn and email.
We run our own outbound with an agent on autopilot. When it makes a mistake, we write down the rule it was missing, together with the number that showed us the mistake. This repo is that rulebook, with our client data removed.
A few of the numbers inside:
- Free LinkedIn accounts, invite with a note: 66% refused. Same accounts, no note: 0.3%.
- A 30-day "closing the loop" message went out 133 times and got 2 replies.
- We banned "curious" from our copy. Ten days later "worth 15 minutes" had appeared 58 times.
- 630 of 692 messages went out as one block of text with no line breaks.
- A dedup check reported 2% duplicates. The emails people actually received had 48%.
What's inside
skills/
├── targeting-* 5 who you write to, and why this week
├── first-touch-* 5 the invite, the note, the first message
├── writing-* 5 what the text reads like to a human
├── replies-* 4 triage, live threads, silence
├── sending-* 4 pace, refusals, keeping profiles alive
├── learning-* 4 weekly readout, A/B splits, checks
└── meetings-* 2 brief before, reminder before
scripts/ 3 batch checks the model should not eyeball
| Group | Skill | The rule |
|---|
| targeting | targeting-sales-hiring-signal | A job post is a reason to write only when they hire in sales |
| targeting-fact-with-source | One fact, with its source, or no message |
| targeting-named-person-not-inbox | A shared inbox is not a contact. Find the names yourself |
| targeting-reason-can-be-context | The reason to write can be context. Zero events is a valid answer |
| targeting-exclude-existing-customers | Exclude existing customers from their own export, before the first send |
| first-touch | first-touch-straight-to-the-point | Say what you sell in the first message, and keep it short |
| first-touch-invite-note-by-account-type | Invite notes from Premium accounts only |
| first-touch-message-after-accept | The first message after accept states intent and adds something new |
| first-touch-booking-link-once | The booking link goes into a thread once |
| first-touch-no-negation | Say it as a statement. Name no conflict |
| writing | writing-stamp-is-frequency | A cliche is measured by how often it repeats across a batch |
| writing-paragraphs-not-walls | Paragraphs, with a blank line between them |
| writing-de-ai-final-pass | Final pass: make it read like one person wrote it |
| writing-weekday-matches-date | A weekday next to a date is checked against the calendar |
| writing-claim-needs-ground | Every claim travels with its ground |
| replies | replies-triage | Sort every reply before anyone answers it |
| replies-discovery-not-pitch | A live thread is questions first, product second |
| replies-follow-up-after-silence | After silence, ask about timing. Skip the goodbye |
| replies-48h-sla | No lead waits more than 48 hours when the ball is ours |
| sending | sending-spread-the-day | Spread the day across the working window |
| sending-refusal-costs-an-attempt | A refusal costs one attempt, not a day |
| sending-protect-real-profiles | Mechanics experiments run on disposable profiles only |
| sending-limits-from-vendor-docs | Limits come from the docs and your own measurements |
| learning | learning-weekly-readout | Weekly readout: one change, and only on enough data |
| learning-decision-needs-why | Every change logs why, what you expect, and what you rejected |
| learning-ab-split-key | Split by lead, lock the arm, keep the key smaller than a batch |
| learning-check-what-the-human-received | Check what the human received |
| meetings | meetings-pre-call-brief | A brief before every external call |
| meetings-reminder-hour-before | One short reminder, one hour before, in the channel you already use |
Scripts (Python 3.9+, standard library only):
scripts/sameness.py checks a batch for duplicates, repeated phrases and repeated endings.scripts/paragraphs.py splits a message into paragraphs, or --check fails on a wall of text.scripts/weekday_check.py fails when a weekday and a date disagree ("Monday, September 22").
Install
Claude Code, as a plugin:
/plugin marketplace add dostrovcev/outbound-agent-skills
/plugin install outbound-agent-skills@revault
Or copy the folders you want into ~/.claude/skills/ (all projects) or .claude/skills/ (one project). Claude picks a skill up from its description when the task matches.
In Claude.ai or the Claude app: zip a single skill folder and upload it under Settings, Capabilities, Skills.
How to use them
Each skill has the same shape: the rule, why (with our number), steps, a pre-send checklist, and an example. If you pick two, start with writing-stamp-is-frequency and first-touch-invite-note-by-account-type: both work on the next batch you send.
The numbers come from our own sending across several LinkedIn profiles in 2026. Your market will give you different numbers. Keep the method, measure your own, and replace ours.
The engine behind the rules
These rules run inside Re:Vault, an outbound engine: sourcing, sending across LinkedIn profiles, reply handling and a weekly readout. You can connect your own Claude to it and run it yourself, from $199 a month: re-vault.io/mcp.
License
MIT. Use it, fork it, change the numbers to yours. If a rule saves you a profile, tell us which one.
