Work / AI systems

Builder Funnel + my own work · 2026

An agent team for the work nobody wants to do.

An agency runs on repeat work: audits, weekly reports, prep before every sales call. I turned the parts that follow the same steps every time into Claude Code skills and agent pipelines, and kept the judgment calls for myself.

Claude CodeAI agentsAPIsAutomation
30+custom skills in weekly use across reporting, research, audits and content
4sales-call deliverables built from one command for each prospect
75past strategy plans turned into a reusable repo the strategy agents read first

TL;DR

  • 30+ custom Claude Code skills I use every week, each with one narrow job and a written playbook.
  • The big ones run as pipelines: one command pulls the data, several agents do their piece, and the output lands on a review page.
  • Nothing goes to a client, an ad account or a prospect until a person has looked at it.

The problem

A lot of agency work follows the same steps every time. Pull the numbers, check the site, read the call notes, fill in the template. It's important work, and it eats the hours that should go to strategy.

Chat tools helped a little, but every session started from zero. I wanted systems that remember how we do things and can reach our real tools.

What I built

  1. Sales-call prep. Before a framework call, one command pulls the deal from HubSpot and the discovery call transcript, audits the prospect's site, picks a matching case study, then builds the intake form, a filled scorecard, a growth snapshot PDF and the call deck.
  2. Blueprint audits. For a new client, an agent team pulls search data, crawls the site, checks competitors, directories, social profiles and AI search visibility, then writes the raw audit files the strategy draws from.
  3. Weekly ads reporting. Pulls new leads, flags ones tagged "direct traffic" that probably came from an ad, checks stage movement, runs the funnel math against spend and posts the report to the project board.
  4. Voice of customer. Pulls sales calls from the call recorder, pulls out objections and the words buyers use, and publishes them to the company dashboard.
In
  • HubSpot deal + contact
  • Discovery call transcript
  • Prospect's website
  • Search data
Agents
  • Site + SEO audit
  • Case study match
  • Scorecard fill
  • Deck + snapshot build
Out
  • Intake form
  • Filled scorecard
  • Growth snapshot PDF
  • Framework call deck
ThenI review it↑ Nothing ships without this step
The sales-call prep pipeline. Each agent gets one narrow job, and the output waits on a review page.

How I keep it honest

Every pipeline ends on a review page. Ad creative gets approved before anything uploads to Meta. Blueprint copy gets approved in markdown before it goes into the client deck. I even built a markup layer for those pages, so I can circle things and drop notes for the agent instead of typing out where the problem is.

Narrow beats clever. The skills that stuck do one thing and have a written playbook. The ones that tried to do everything needed babysitting.

I also wrote a 20-skill Claude starter kit for the team and trained teammates on it, since not everyone works in a terminal.

Outside of client work

The same setup runs my own workflows: a daily job board scan, LinkedIn content batching, prospecting for my Kickstarter side business, and the editing pipeline for a podcast I co-host.

What I'd do differently

I'd write the review step first. My early skills went straight to output, and fixing a confident mistake after the fact costs more than catching it on a review page.

I'd also split big skills sooner. When one skill starts handling three jobs, it's time for three skills.

Built withClaude CodeHubSpot APIMeta + Google Ads APIsAhrefsPlaywrightPython