Watch the full video: Stop Building Generic Clay Tables. Make Claude Copy Your Taste.
Tables used to cap at 50,000 rows. That one number is why most Clay tutorials still teach you to build a one-off sheet, enrich it, export it, and start over next month. We stopped doing that.
I run Growth Engine X. We send 8 million emails a month. I am Clay's largest user. The 2024 beginner walkthrough I recorded (101 Tutorial for Clay.com) still holds for columns, waterfalls, and credit hygiene. What changed is the job. Clay is no longer a spreadsheet you click through. It is a function library you call from Claude, plus an Audience that keeps your TAM alive after you close the laptop.
This is the operator version. Three moves: make Claude copy your Clay taste, treat tables as functions, and run Audiences as a living TAM. If you want us to gift the first campaign instead of watching you rebuild this for two weeks, the free test is at coldoutbound.com.
The Complaint I Had After One Day on the CLI
Clay shipped the CLI/API so you can build workflows straight from Claude. I had been stress-testing it with Kareem, Matthew Quan, and Ryan Chang since the March demo. After one real day, the complaint was not "it cannot build a table." It was taste.
"The Clay/Claude integration would choose to use integrations and AI models I wouldn't pick."
Off the shelf, Claude will pick a Google search provider I do not use. It will pick a model I do not run. It will write prompts that do not cache. We always want OpenWeb Ninja for search, 4o mini for the cheap AI columns, and prompts structured so the examples stay in the cache. Pulling the CLI without that context is how you get a table that "works" and still burns credits like a junior SDR.
The fix is one prompt. Because the CLI can read your existing tables, you do not write a style guide from memory. You make it extract one.
"Review my most recent clay tables {{inject your time horizon you want here}}. Analyze the structure and the decisions that I make when building my Clay workflows that are in my account. Then create your own rule book that gets saved to this skill so you can keep building workflows the same way I build tables. Output your top 10 rules and show me so I can confirm them."
That is the whole Clay tutorial now. Not "add a column, pick an enrichment, map the field." Teach the agent the ten rules you already follow, confirm them, save them as a skill, then let it build.
Tables as Functions, Not Art Projects
Matthew Quan gave us early access to Clay functions that run inside Claude and Codex. The first function we promoted was our internal email waterfall. Order, nothing accepted without LeadMagic valid:
- QuickEnrich
- BlitzAPI
- Prospeo (Bounceban)
- Icypeas
- SIGNALIZ
- Smartlead Email Finder last
Source: Matthew Quan early access post
Most Clay tables are combinations of functions. Before functions, we had the same waterfall baked into 40+ tables. Change the order once, open 40 tables. Now we change the function once and every table that calls it updates.
We pair this with Auto Clay + goal mode in Claude Code. The agent does not rebuild the waterfall from scratch every time. It calls the function, checks the output, and moves on. Our skills repo is public: coldoutboundskills on GitHub.
If you are still learning the column-level basics, start with the original walkthrough videos:
- 101 Tutorial for Clay.com — columns, waterfalls, credit hygiene
- Advanced Clay workflows — scaling past beginner tables
- Clay automation deep dive — scheduled sources and evergreen campaigns
The shift is mental: a table is not a deliverable. A table is a function you can call, version, and reuse.
Browser Use: When Screenshots Are Not Enough
Most browser automation works off screenshots. You click, you wait, you screenshot, you guess. browser-use also reads the front-end API.
"Most browser automation works off screenshots... browser-use also reads the front-end API."
That matters when you are scraping sites that do not expose clean APIs. We use browser-use inside Claude Code workflows when a standard HTTP fetch will not get the data. See the walkthrough: browser-use demo.
The pattern: let Claude build the Clay table structure, call browser-use for the hard scrape, push results back into Clay for enrichment. You are not clicking through the UI. You are orchestrating.
Audiences Are a Living TAM
The 50,000-row table limit was the old ceiling. Audiences have no row limit and auto-update.
"Audiences don't have the 50k row limit and they auto-update."
We run a living new-in-role list at coldoutbound.com/lists/newinrole. It is not a CSV we re-upload every Monday. It is an Audience that refreshes as people change jobs.
The workflow:
- Build your ICP filters once in Clay
- Promote the table to an Audience
- Connect the Audience to your sequencer or CRM
- Let it run
More on the Audience setup: living TAM post.
If you are still exporting CSVs from Clay and uploading to Smartlead every week, you are doing 2024 Clay. Audiences are 2026 Clay.
What I Would Do This Week
If I were starting from scratch today:
- Run the rule-book prompt — let Claude read your last 30 days of tables and extract your top 10 rules. Confirm them. Save as a skill.
- Promote your best waterfall to a function — email validation, CRM sync, whatever you repeat across tables. One source of truth.
- Launch one living Audience — pick your highest-ROI segment (new in role, headcount growth, whatever signal you trust) and stop re-building the list.
For the full Claude Code loop that runs campaigns without you writing prompts every time, see The Claude Code Loop That Runs Cold Email.
If you want us to run the first campaign instead of watching you rebuild this for two weeks, apply for the free test at coldoutbound.com.
Key Takeaways
- Taste is the bottleneck — the CLI builds tables, but you need to teach it your integration and model preferences first
- One prompt extracts your rule book — Claude reads your existing tables and codifies your decisions into a reusable skill
- Tables are functions — promote waterfalls to Clay functions so you change logic once, not in 40 tables
- Audiences replace CSV exports — no row limit, auto-update, living TAM
- browser-use reads front-end APIs — use it when screenshot-based automation fails
- Skills repo: github.com/growthenginenowoslawski/coldoutboundskills