The Campaign Launcher is a skill for Claude that takes a marketing experiment from ICP and positioning input to live campaigns: it drafts the plan, produces the assets, configures Google Ads, Meta Ads, and email outreach in parallel, and holds at two human gates before any money moves. In March I published a LinkedIn post about it and asked people to comment LAUNCHER to receive the files. The thread ran to 743 comments. This article is the other half of that promise: every file that was behind the gate, linked directly. No comment, no email, no DM required.

This expands on the LinkedIn post I published in March about the Campaign Launcher, which became the most commented post I have written.

Key Takeaways

  • The Campaign Launcher skill, its setup guide, and all eight reference files are public on GitHub. Every link is in this article. Nothing in it is gated.
  • The skill orchestrates four phases: research and planning, asset generation, channel configuration, and activation with post-launch monitoring.
  • Two human gates are built in: you review the plan before any budget is committed, and you explicitly confirm activation before campaigns go live.
  • The open version runs standalone on your own API keys in Claude Code or Claude Desktop. It does not require Improvado.
  • A skill automates the workflow. It inherits whatever your data layer knows about your accounts, naming, and results. That layer is the part a skill cannot give you.

What Was Behind the Gate

The March post promised three things to commenters: the skill itself, a step-by-step guide, and instructions for running it in Claude Desktop or Claude Code. All three live in our public repository, and here they are:

  • SKILL.md: the skill definition itself. This is the file Claude reads. It contains the full pipeline, the phase architecture, and the gate logic.
  • Setup guide: prerequisites, API keys, and the install path for Claude Code and Claude Desktop.
  • The reference library: eight files the skill loads on demand. The experiment plan template, the copy generator, the creative generator, one playbook per channel for Google Ads, Meta Ads, and email outreach, and an integration note for teams that do run Improvado.

Why gate it in the first place? A comment gate is a distribution experiment: it trades friction for reach, because every comment pushes the post to more feeds. The experiment worked as distribution. But a gate is the wrong permanent home for a working artifact, and delivering files one DM at a time does not scale. A repository link in a blog post does. So this is where it lives now.

What the Skill Actually Does

The Campaign Launcher is not a prompt. It is a workflow definition with a prerequisites check, four working phases, and two mandatory stops. When you invoke it with an ICP and a positioning statement, this is what runs:

Phase 1: Research and draft plan

Claude researches the audience and the channels, then writes a full experiment plan into a file you can edit: hypothesis, budget split across channels, audience and persona definitions, duration, and the target KPI and CPA. In the March post I claimed this compresses about three days of planning work into minutes. The plan is a markdown file, not a black box. Change any line before approving it.

Review gate

Nothing proceeds until you approve the plan, in chat or by editing the file. No budget is committed at this point. This is the stop that makes the speed safe.

Phase 2: Asset generation

Copy variants and creative concepts for every channel in the plan, generated in parallel against the copy and creative reference files. The March post cited twenty ad variations in about seven minutes. Treat numbers like that as what they are: illustrations of where the time goes, not a guarantee. The honest claim is that asset generation stops being the bottleneck.

Phase 3: Channel configuration

Campaign structures for Google Ads and Meta Ads and sequences for email outreach, built with your own keys through each channel's API, or through CSV export where you would rather upload than connect. Hours of clicking through three ad managers becomes configuration Claude writes and you inspect.

Activation gate and Phase 4

The second stop: campaigns exist but are not live until you explicitly confirm activation, per channel or all at once. Phase 4 then activates what you confirmed and carries the post-launch follow-up inside the same conversation.

How to Run It Yourself

The short version, with the full detail in the setup guide:

  • Clone the repository and install the skill where your Claude can load it. The repository ships as a plugin with a manifest, and the setup guide walks through the Claude Code path.
  • Bring your own API keys for the channels you want: Google Ads, Meta, and your email tool. The open version has no dependency on our infrastructure.
  • Expect a real conversation before the first plan. On first run the skill insists on a business discovery pass: ten blocks covering your product, customers, buying process, past campaigns, and competitors, one block at a time. It takes about ten minutes, and it is the difference between a plan built on your business and a plan built on general knowledge.

Run it on a small test budget first. The gates protect you from the obvious failure mode, but the plan is only as good as the inputs you give it, and the discovery conversation is where you find out what your inputs are missing.

Talk to an expert if you want to see what this workflow looks like when it is wired to governed data instead of raw API keys.

Where a Skill Stops and Infrastructure Begins

Here is the part the March post compressed into a P.S., and it deserves more than a P.S.

A skill file automates a workflow. It does not know your historical CPL by segment, which naming convention your ad accounts actually follow, or what happened the last time you ran this audience. When Claude plans a campaign with no data layer underneath, it plans from general knowledge, and general knowledge is what your competitors also have. The interesting version of this workflow is the one where Phase 1 reads your real performance history and Phase 4 writes results back somewhere they will be found next quarter. That is not a skill problem. That is a data foundation problem, and it is the same problem whether a human or an agent runs the campaign. I wrote about the organizational side of this in AI adoption as an org design question, and about what agents without memory cost you in the hidden costs of AI agents.

If you want the broader map of what skills can and cannot carry in a marketing stack, the team maintains a full guide to building Claude skills for marketing. The Campaign Launcher is one entry in that landscape, and the guide covers testing, validation, and the infrastructure requirements honestly.

Talk to an expert about connecting a campaign workflow like this to a governed marketing data foundation.

Frequently Asked Questions

Is the Campaign Launcher skill free?

The repository is public on GitHub and the links in this article open without any form or payment. You pay for your own Claude usage and your own ad spend, nothing to us.

Does it require Improvado?

No. The open version runs standalone on your own API keys. The repository includes an integration reference for teams that do use Improvado, but it is optional.

Which channels does it support?

Google Ads, Meta Ads, and email outreach, each with its own reference playbook. The architecture is per-channel files, so adding a channel means adding a reference file, not rewriting the skill.

Can it spend money without my approval?

No. Two human gates are built into the pipeline: the plan review before any budget is committed, and an explicit activation confirmation before campaigns go live.

Why is it no longer gated behind a comment?

The comment gate was a distribution experiment on one LinkedIn post, and it did its job there. A public repository linked from a blog post is a better permanent home for a working artifact than a DM queue.

What is the difference between this and just prompting Claude to plan a campaign?

Structure and stops. The skill enforces a fixed pipeline with reference playbooks per phase and two mandatory human gates, which makes runs repeatable and reviewable. A freeform prompt gives you a plan; a skill gives you a process.