Using Claude (and Grok) for Outbound Research — Without Throwing Out Clay on Day One
The stack argument has a new headline. People post "CLAY OR CLAUDE CODE?" as if you had to pick a parent. Others say they stopped paying for Clay, and that Claude Code now builds all their outbound campaigns. A third flavour asks whether you want to build an entire AI sales team with 50 Claude Code skills.
Pramod Kumar, co-founder of Kingdom of Kumar. On X
·
I use models for this work. Kingdom of Kumar's systems run on Grok and Claude. I still will not treat a model, or a folder of skills, as a sales team, and I will not tell you to cancel Clay on the strength of a thread. This is how I would use Claude — and Grok — for outbound research, next to a workbench if you already have one, behind an approval rule if a first line will be sent.
Related: When not to buy Clay, An AI SDR approval workflow, Cold email personalization that isn't fake. The work is on GTM engineer and GTM engineering. If the product is not ready, take the App Launch Score first.
I will not run a model bake-off in this article, and I will not invent win rates. Use the model you will actually review, and keep evidence in a field you own.
"CLAY OR CLAUDE CODE?"
It is usually the wrong question. They are not the same job.
Clay, or a simpler set of APIs, is for getting and shaping data: emails, firmographics, a waterfall, a table you can cap. Claude or Grok is for reading, summarising, drafting, and helping you think — if you give them sources, and if a person checks claims. You can do research in a workbench, and you can do some data work with an agent, but if you force an either/or, you will either overpay for a workbench you use as a chat window, or you will ask a model to invent contact data it does not have.
If you already pay for Clay and it earns a specific step, keep it while you add a model as a copilot for research and drafts. If you do not pay for Clay, you do not have to buy it to use a model. See the "when not to buy" piece.
"We stopped paying for Clay… Now Claude Code builds all our outbound campaigns"
Sometimes that is a careful, cheaper path: an operator who knows the motion, APIs for data, a model for research and drafts, a sequencer or mailbox they still control. Sometimes it is a prompt on a list with no verification and no approval. The sentence alone cannot tell you which.
If you want to leave Clay, export, write down the steps you still trust, and keep verification and approval. If "Claude Code builds the campaigns" means a person still reviews first lines and the company still owns the workspace, I can live with that description. If it means an agent hits send, I cannot.
Claude Code, specifically, is a way of working that suits people who will supervise an agent in a project. It is not automatically the right interface for a non-technical founder. A chat with Claude or Grok, plus a sheet, can be the same idea at a gentler slope.
"Want to build an entire AI sales team with 50 Claude Code skills?"
I would not. A pile of skills is a pile of prompts and tools. It is not a team. Teams have judgement, a fit rule, someone who answers a sceptical reply, and someone who stops a bad batch. Fifty skills is also a maintenance problem. I would rather have three tightly scoped jobs — research this account from allowed sources, draft a first line from one evidence field, summarise the Friday note — than a simulated department.
If a post is really selling a pack of skills, ask to see the approval surface and the last ten rejected drafts. If there are no rejects, there is no standard.
A research workflow I would actually use
- 1The account is already a fit, or it is being checked for fit — the model does not get to invent the ICP.
- 2You collect, or the model is pointed at, a small set of sources you are allowed to use: a site page, a post, a job listing, notes from a past call. You do not ask it to "find an email and guess their budget."
- 3The output is structured: facts with sources, facts you could not verify, and a suggested first line that uses only the verified facts — or a generic line if the facts are thin.
- 4A person accepts, edits, or throws out the first line. See the personalization and approval articles.
- 5Evidence is stored on the record, dated.
Grok or Claude can both sit in step 3. I will not tell you which is "better for GTM" as a slogan. I will tell you to keep the same constraints for both.
What I would not use a model for, in this motion
- Inventing contact details
- Declaring buying intent from a vibe
- Writing a first line that mentions geography, funding, or a hire you did not supply
- Running the live reply thread without you
- Replacing the Friday review
Those limits are the difference between a research assistant and a liability.
A one-hour test you can run on ten accounts
Pick ten companies you already believe are a fit. For each, give the model only two or three sources you have opened yourself, and ask for the structured output above. Then, without the model, glance at the same sources and mark: true, invented, or over-read.
You are not scoring the brand of the model. You are scoring your workflow. If inventions show up, tighten the prompt and the source set before you ever connect this to a sequencer. If the drafts are mostly generic because the sources are thin, that is a useful result too: you do not have a personalization fact, and you should not pretend otherwise.
Do this test even if you already like the model's voice. Voice is not the failure mode. Confident, unsourced detail is.
Working in the company's workspace
Whether you use a chat, Claude Code, or another agent interface, the notes, prompts, and evidence should live in a place the company owns. A freelancer's local project with no export is a risk, the same as a personal Clay workspace. If you leave a tool, you should still have the `why_today` sentence and the source you used.
FAQ
Should I cancel Clay to "move to Claude"?
Only after an audit of what Clay is actually doing, and a plan for data and verification. A model is not a waterfall.
Is Claude Code required?
No. It can be excellent for an operator who works that way. A founder can get the value of research-plus-draft from a normal chat, with the same evidence rules.
Is this what the $500/month offer is?
Kingdom of Kumar builds and runs a scoped system on Grok and Claude for $500 a month, three-month minimum. You own it, and you approve what goes out. We will use the models as copilots, not as a fifty-skill sales team, and we will not throw out a useful workbench for a headline. This is not a $135,000 hire.