I built an agent to reverse-engineer how great marketing actually works

I’ve been on a journey to transform repeated tasks in marketing to composable agents and skills.

(If you want to see the entire kit and caboodle, everything is on my Github: https://github.com/j1ngg/)

A high leverage task for agents is competitive marketing research. The old way is to search everywhere (Google, LinkedIn, X, Reddit, YouTube, Hacker News…) collect, sort, and make sense of it.

The problem with that is that it’s:

1. Time consuming

2. Easy to miss hidden channels

3. Difficult to see the distribution model together from channel-specific patchwork

We can do better! Enter a shiny new Claude agent: how-they-market

It’s an agent that researches how companies and creators market themselves and generates the marketing model aka the actual system underneath the visible activity.

The goal is simple: it helps you get from “huh, they seem good” to “this is exactly what I should test” much faster.

What how-they-market helps you do#

If you work in marketing, you’ve probably had this experience:

You see a company or creator that is obviously doing something right. They keep showing up. Their content travels. Their positioning feels sharp. Their launch gets traction.

And when you try to study them, you end up with a messy pile of links and observations. Like ok they:

• Post on LinkedIn

• Have a founder that is active on X

• Have a newsletter

• Launched on Product Hunt

Cool. But it doesn’t tell you which channel is doing discovery work. It doesn’t tell you what builds trust. It doesn’t tell you what is likely converting. It doesn’t tell you what is core to the strategy versus what is just aesthetic spillover.

That’s the gap this agent is meant to close.

Instead of giving you a prettier pile of research, it tries to help you answer the only questions that really matter:

In plain English, this agent helps you:

• How does the influencer or competitor’s marketing system work?

• Why is it effective for turning attention into trust and conversion?

• Which parts are transferable?

• What should you try in your own context?

What the agent actually does#

At a high level, the agent researches how a company or creator markets themselves and returns a structured analysis of their marketing strategy.

The workflow is pretty straightforward:

1. Discovery: broad search depending on whether the subject is a company or a creator

2. Primary source research: full-source retrieval instead of relying on snippets

3. Follow-up exploration: deeper digging where there are actual gaps or promising threads

Then it organizes the findings around the thing that matters most:

The marketing distribution model

Meaning: how distribution, content, trust, and positioning fit together into a coherent system.

It also breaks out:

• Channels

• Content strategy

• Monetization

• Evolution over time

• Voice and tone

• Community presence

• Key tactics

• What to steal

The goal is not to hand you a pretty artifact about someone else’s brilliance. It’s to hand you something you can actually use.

How to use it#

1. Open Terminal (on a Mac it’s Command + Space and then search for Terminal)

2. Install Claude Code (https://code.claude.com/docs/en/quickstart)

   1. P.S. Trust yourself to do this! The command line can look intimidating but it’s not difficult and it’ll unlock a universe of capabilities.

3. Run Claude Code by typing claude and pressing enter

4. Then in natural language you can tell Claude to help you try this agent: https://github.com/j1ngg/how-they-market

5. Enter the person or company you want to research

6. The agent will run and return a completed report saved to your computer. If you can’t find it, ask Claude!

Also, Claude will ask if it can do web search, run git commands, etc. You should give it permissions.

What was hard about building it#

As always, the build is never quite as straightforward as one might believe.

At first the output was too shallow. It didn’t cite well. It wasn’t better than just asking Claude in chat.

Then, after more iteration, it swung too far in the other direction. It got more detailed, but also heavier and less usable. Which is its own kind of failure.

That was probably the biggest lesson from the build:

Better AI workflows are not just about adding capability. They’re about enforcing judgment of what’s important to the end-user.

What should the tool ignore? Where should it go deep? What belongs in the final output versus the research trace? How do you make the result useful enough that someone can act on it quickly?

If you want to build agents yourself, that’s probably the most practical takeaway here: don’t start by asking how to make the agent do more. Start by asking how to make the output more useful.

A capable agent that produces bloated, vague output is still a bad tool. A narrower agent with better judgment is usually much more valuable.

How this fits into a bigger GTM workflow#

The agent on its own is useful. But the more interesting move is embedding it inside a bigger workflow.

I’ve started using it inside another repo that walks the user through a messaging and positioning workshop to set the foundations of a launch (I’ll do a write up later but here’s the repo if you want to play with it, try the skill messaging and positioning first).

That’s where the value of composable agents starts to show up.

Instead of treating research like a standalone artifact, the competitive marketing research agent becomes one step in a chain. It generates strategic input that then feeds into a more guided workflow around positioning, messaging, and launch prep.

So instead of stopping at “here’s how this company markets,” the broader system can help shape launch strategy, distribution channels, experimentation loops, even messaging angles that have a better shot of getting through. No more posting into the void for a launch and hoping something happens from it (yay).

Agent chaining to unlock broader, more contextualized, repeatable and scalable workflows is SO much more exciting than a giant magical prompt with some tools and loops (aka an agent). Essentially you end up with a set of tools that each do a specific job well and then hand off into the next stage of the work.

And I think there’s a lot more room to build useful systems like that.

More on that soon.