An outbound skill for Claude Code is a packaged set of instructions, saved as a SKILL.md file, that teaches Claude Code your outreach rules once (your ICP, your voice, your sequence structure) so it applies them consistently every time you ask it to research an account or draft a message, instead of you re-explaining the brief in every session.
This guide is for founders, sales leaders, RevOps, and SDRs who want Claude Code to help with prospecting and outreach and are deciding whether to write their own skill, install a pre-built one, or both. It covers what a skill actually is under the hood, how to write one, and the one thing a skill cannot do for you no matter how well it is written.
What is an outbound skill for Claude Code?
It is a folder (.claude/skills/<name>/ for one project, ~/.claude/skills/<name>/ for every project you work in) containing at least a SKILL.md file: a short frontmatter block plus plain-English instructions that Claude Code reads and follows when the task matches.
For outbound work specifically, that usually means codifying things like:
- How to describe your ideal customer profile (ICP) so Claude applies it consistently.
- The structure of a cold email or LinkedIn message in your voice (length, opening line, call to action).
- What a "qualified" account looks like before it is worth drafting outreach for.
- How to score or prioritize a list of accounts once you have one.
Several outbound skills already exist as public repositories you can install rather than write from scratch, bundling things like persona research, cold email drafting, and pipeline analysis into one package. Installing one is a reasonable starting point. The rest of this guide covers what is happening underneath so you can adapt one, or write your own when the pre-built version does not match how your team actually sells.
How does a Claude Code skill actually work?
A skill is not code that runs; it is an instruction pack that Claude Code reads and reasons from. Two things make that work:
- Frontmatter with a
nameand adescription. Thedescriptionfield is what Claude Code scans across all your installed skills to decide which one applies to the current request. A vague description ("helps with sales") gets skipped in favor of a more specific one; a description that says exactly what the skill does and when to use it gets picked up reliably. - The instructions themselves. Everything below the frontmatter is plain text: steps, rules, examples, the format you want output in. Claude Code follows it the way it would follow a detailed prompt, except the prompt is saved and reused instead of retyped.
When a session starts, Claude Code reads the descriptions of the skills available to it. As you work, it checks whether your current request matches one of those descriptions and applies the matching skill automatically, or you can name it explicitly.
How do you write an outbound skill, step by step?
- Pick one job per skill. A skill that tries to cover ICP definition, email drafting, and pipeline analysis at once gets vague instructions that match too many requests. Separate skills (or clearly separated sections) for "define the ICP," "draft the first-touch message," and "qualify this account" each stay sharp.
- Write the
descriptionfirst. State exactly what the skill does and when it should trigger, for example: "Drafts a first-touch cold email for a specific account once its ICP fit and a buying signal are known. Use when asked to write outreach for a named company." That specificity is what keeps Claude Code from firing the skill on the wrong task. - Give it your real constraints, not generic sales advice. Claude already knows the generic structure of a cold email. What it does not know is your voice, your length limit, which claims you are allowed to make, and which ones legal or product has ruled out. Put those in.
- Include a worked example. One real before/after, or one filled-in template, teaches the format faster than a paragraph of rules describing it.
- Separate the instructions from the data. The skill should say how to use an ICP and a signal once it has them. It should not contain the ICP or a list of target accounts hardcoded in the file, because that list goes stale and has to be maintained by hand everywhere the skill is copied.
- Test it on a real account, not a hypothetical one, and tighten the instructions wherever the output misses your actual bar.
A minimal example for the "qualify this account" step:
---
name: qualify-account
description: Checks whether a named company matches our ICP and has a live buying
signal worth acting on. Use when asked whether a specific company is worth
prospecting.
---
Given a company name:
1. Check it against our ICP: B2B SaaS, 50-500 employees, US or UK HQ.
2. Look for a buying signal from the last 90 days: new funding round, a
relevant open role, or a leadership change in a buying-committee title.
3. If both are true, output: company name, one-line reason it fits, the
signal and its source. If either is missing, say which one and stop;
do not draft outreach for an unqualified account.
What should an outbound skill not try to do on its own?
A skill is an instruction layer, not a data source. It tells Claude Code how to use an ICP and a signal once it has them; it does not give Claude Code a live, current view of which companies exist, which ones match your ICP today, or which ones just did something worth noticing. Without a connection to a real, continuously updated company dataset, Claude Code is reasoning from its training data when asked "find me 50 companies like this," and training data does not know who was hired last week or who raised a round last month.
This is the gap that shows up most often: a nicely written skill produces a nicely formatted, confidently wrong list, because the instructions were sound but the input was Claude guessing rather than Claude looking something up. The fix is not a better-written skill; it is connecting the skill to a tool that can answer "which real companies match this ICP, right now" before the skill's drafting instructions ever run.
That same input problem shows up in narrower prospecting tasks too. If your ICP depends on social activity rather than firmographics alone, see which platforms combine social engagement data with ICP matching for what that combination actually requires. And if your qualification criteria include a workforce signal like contingent staffing use, how to find companies that use staffing agencies covers where that kind of signal is findable at all before you write a skill that assumes it.
How does Retriever fit into an outbound skill workflow?
Retriever lets you build AI agents for lead generation, and they cover the exact step a skill cannot do on its own: finding and verifying which companies currently match. Instead of filtering a static database, your agent reads the live web (a company's website, its job posts, its reports), and every lead is checked against your ICP with the source behind it, so the list going into your skill is checkable rather than assumed.
Buying signals are tracked as part of the same agent instead of bolted on afterward. Retriever sets up and continuously monitors hiring, funding, tech-stack, and leadership signals, plus LinkedIn and social listening, so a company surfaces when it fits your ICP and carries a signal worth acting on, which is exactly the "qualify this account" step above.
Retriever also runs an MCP server, so an outbound skill in Claude Code can call Retriever's data and agents directly as a tool rather than asking Claude to guess. Retriever supplies the verified, current list and the evidence behind it; your skill's instructions handle the formatting, voice, and sequencing on top. Book a demo or see what data and signals agents use to connect it to your own ICP.
Frequently asked questions
Do I need to write my own outbound skill, or can I install one? Installing a pre-built outbound skill is a reasonable starting point, and several public repositories bundle things like ICP definition, cold email drafting, and pipeline analysis. Write your own, or adapt an installed one, once the generic version does not match your actual voice, qualification bar, or sequence structure.
Does a Claude Code skill run automatically, or do I have to trigger it? Both are possible. Claude Code reads each installed skill's description and can apply a matching skill automatically when your request fits it, or you can name the skill explicitly if you want to be sure which one runs.
Can an outbound skill find new accounts on its own? Not reliably by itself. The skill's instructions tell Claude Code how to evaluate or draft for an account once it has one; finding which real companies currently match your ICP requires connecting Claude Code to a live company-data source, typically through MCP, rather than relying on the model's general knowledge.
What is the difference between a skill and an MCP server? A skill is a set of instructions Claude Code follows; an MCP server is a tool Claude Code can call to get real data or take an action. In an outbound workflow they work together: the skill defines how to qualify and draft, and an MCP-connected tool supplies the current company and signal data the skill reasons over.
Where should I store an outbound skill if my whole team uses it?
In the project's .claude/skills/<name>/ folder if it is specific to one codebase or workspace, or in a shared, version-controlled location your team installs from if everyone should use the same ICP and voice rules. A skill that only lives on one person's machine drifts from what everyone else is sending.