Lieben runs cold email campaigns in Claude Code, not Clay
Ten enrichment agents, five reviewer agents and a hard stop before launch, with the reply count coming from one demo list.
Michel Lieben published a working build of an outbound campaign run inside Claude Code instead of Clay. He says his Head of Service Delivery, Soheil, runs four campaigns at once in four terminals, dictating prompts with Wispr Flow. One campaign started with 6,000 trade show names, ended with a little over 1,000 validated addresses, and got 19 positive replies in its first week. Clay never opened.
How the run is structured
The sequence starts in plan mode, where the agent reports which files it read, how it intends to find missing domains, and which rows it will process, before it spends anything. Then the rows get split. The prompt is "Spin up 10 agents and split the remaining rows between them." Lieben says the enrichment calls do not change when they are split ten ways, so credit spend holds steady and the job finishes faster.
Two cost controls matter more than the parallelism. The waterfall only runs on Tier 1 and Tier 2, because every tier waved through spends credits finding emails for people you were not going to write to first. Phone lookups stay on Tier 1 only, since Lieben puts a phone lookup at around ten times the cost of an email one. When LeadMagic was slow that day, the instruction was to stop waiting on the monitor and build the list from the domains already in hand.
Five reviewers before a human reads it
Copy is drafted by a copywriter skill, read by five reviewer agents, and rewritten against their notes until it scores at least 7 out of 10. Nobody on the team sees it before that. The reviewer Lieben would build first is the prospect, pointed at an ICP matrix and told to read the draft once, fast, on a phone, then open with a verdict of reply, ignore or delete. On this build it cast itself as the managing director of a family spice export firm reading between WhatsApp messages from buyers.
The upload prompt ends with "Do not set anything live." The agent maps merge fields, loads all three tiers into Instantly and Tier 1 into lemlist for LinkedIn steps, then stops for a human to press launch.
For people sending, the coverage claim is the part to discount. Lieben says a single provider covers around 40% of a list while his stack runs seven finders through the ColdIQ MCP and gets past 80%, and ColdIQ is his own API layer. He also concedes the bottleneck was domain lookup, since most of that trade show export had no website column and he picked spice merchants on purpose as a hard ICP. Treat the 19 replies as one team's demo, not a benchmark.
A single provider covers around 40% of a list. Ours goes through seven finders on the ColdIQ MCP and gets past 80%.