AI AGENTAGENT INFRAENRICHMENT
Agent-Run Client Delivery System
A five-system GTM engagement, run end to end by people who don't code. No API keys. No install. The agent runs the workflow.
Claude Code · Supabase · Granola · Railway · clients/{slug}/state.json
The Problem
A GTM engineering engagement has a shape: understand the ICP, build the TAM, mine the call transcripts for objections and gaps, land it in a database, then test whether the data providers and the detection prompts are actually good. I was the only person who could run it, which meant our capacity was my calendar. So I productized the shape as a Claude Code-native workflow: a thin slash-command router that delegates to six composable skills, with every step writing to a per-client state file so an engagement can pause and resume without losing its place. Non-engineers on the team now deliver it. Seven client folders are live.
Stack
🤖
Claude Code
Slash-command router → 6 composable skills; hooks enforce the workflow
🗃️
Supabase
Client database — accounts, contacts, transcript-mined insights
🎙️
Granola
Call transcripts mined for objections, themes, and gaps
🚀
Railway
Client portal with per-client deliverables sync
📁
clients/{slug}/state.json
Resumable per-client state — pause and pick up anywhere
How It Works
The execution path
Execution flow
Operator runs one slash command— no keys, no install
Router reads clients/{slug}/state.json → resumes at the right step
Skill 1–2: ICP analysis → TAM builder
Skill 3: knowledge hub mines call transcripts for objections / themes / gaps
Skill 4: Supabase database populated; portal syncs deliverables
Skill 5–6: recursive tests benchmark providers and tune ICP-detection prompts
Every step writes state → engagement pauses and resumes without losing place
The design constraint was "no API keys, no Python, no install." If a non-engineer has to configure anything, they won't run it. The agent holds the workflow; the operator holds the judgment. The fifth system is the one I'm proudest of — recursive tests that benchmark data providers against each other and tune the ICP-detection prompts automatically, so the engagement gets better every time it runs.
Key Design Decisions
🚫
No Keys, No Install
The agent runs the workflow. If a non-engineer has to configure anything, adoption is zero. So they configure nothing.
🧩
Thin Router, Fat Skills
One slash command delegates to six composable skills. Each is testable alone; the router only decides what runs next.
⏸️
Resumable by Design
Per-client state.json after every step. Engagements pause for a week and pick up exactly where they stopped.
🔁
Tests That Improve the Engagement
Recursive tests benchmark data providers and tune detection prompts automatically. The system gets better each run.
By The Numbers
0
API keys an operator touches