Results
Built a 17-skill GTM agent system on Claude Code, covering research, competitive monitoring, content, sales ops, and reporting
Shipped five shared context files and five role-agents on a headless, GitHub-first, Slack-integrated architecture, decoupled and Crescendo's to own
Tested every skill on live runs, shipped a clean-installing plugin, and coached Roxanne's team to own and extend the system
Overview
Crescendo runs the contact center for you. Its AI assistants handle voice, chat, email, and shopping; its managed human team takes the cases the AI cannot close; together they resolve around 90% of issues, billed by outcome rather than by seat. Conceived inside General Catalyst in 2023 by operators from Zendesk, Genesys, and Alorica, it raised a $50M Series C at a $500M valuation, and the PartnerHero acquisition gave it a 3,000-person CX workforce. Decagon, Sierra, and Intercom's Fin are the AI-only rivals.
Roxanne, Crescendo's new VP of product marketing, arrived with an architecture already in mind: a headless, GitHub-first, Slack-interfaced agent platform that would run the marketing team's repetitive work on its own. My job, alongside her and the SVP of marketing, Mike Ryan, was to build it and then coach the team to run it. The shift was from doing marketing tasks one at a time to operating a system that does them on a schedule.
Two things made this hard to fake. First, the system had to be decoupled from my machine and my accounts, living on infrastructure Crescendo's own engineers would trust and maintain. Second, it had to survive a handover: install cleanly, pass its tests, and make sense to people who had not built it. A demo would not clear that bar; a shipped, tested, documented system that the team could extend would.
Scope
Think of it as four layers, built bottom up on Claude Code, then handed over with coaching so the team could run and grow it.
The shared context, at the base. I refreshed the ICP across four scored segments, the competitive research on Decagon, Sierra, and Intercom Fin, and the company context, then distilled the load-bearing knowledge into five reusable context files: the lead-scoring rubric, the threat taxonomy, the personas, the products, and the sales narratives. Every skill reads from these, so the whole system speaks with one voice.
17 GTM agent skills on top. They cover the marketing motion end to end: research, competitive monitoring, content, sales ops, reporting. I built all 17 as Claude Code skills, each carrying a worked example so the team sees the output before it runs.
Five role-agents, on a headless architecture. Researcher, growth, paid, content, and sales agents route work to the right skills. The whole platform is GitHub-first and Slack-interfaced: the skills live in a dedicated repo, run without anyone's laptop, report into Slack, and stay decoupled from Genesys, so the system is Crescendo's to own.
Tested, shipped, and coached. Each skill shipped only after passing a live test run, the plugin installs cleanly from the repo, and I handed it over with a quickstart, a skill index, and an executive summary, then ran the office hours that turned the handover into ownership, so Roxanne's team could not only use the skills but write their own and adapt the context as the market moves.
Impact
Crescendo's marketing moved from one-off tasks to an always-on system its own team runs:
17 GTM agent skills, built on Claude Code, covering research, competitive monitoring, content, sales ops, and reporting
Every skill tested on a live run before shipping, with a plugin that installs cleanly from the repo
Five shared context files (lead-scoring rubric, threat taxonomy, personas, products, sales narratives), so every skill speaks with one voice
Five role-agents on a headless, GitHub-first, Slack-integrated architecture, decoupled from Genesys and Crescendo's to own
A team coached to own and extend it: office hours, a quickstart, and an executive summary, so a new VP can run the system and add to it without me in the room
