I Built 99 Claude Code Skills for RevenueFlow. Only 24 Touch Revenue.
Of 99 Claude Code skills I built for RevenueFlow, 24 touch revenue. Here are those 24, grouped by GTM stage, and the three ways this approach breaks.
A Claude Code skill is a folder containing a SKILL.md file that describes a process in plain English, so an agent runs it identically every time. Of 99 skills built at RevenueFlow, 24 touch revenue across six stages: ICP and sourcing, inbox infrastructure, campaign copy, LinkedIn, pipeline, and retention. The remaining 75 are admin and engineering.
Key takeaways
- Of 99 Claude Code skills built for RevenueFlow, 24 touch revenue and 75 are admin and engineering work.
- A skill is a folder with a SKILL.md file holding the process, guardrails, and approval gates, called like a command so it runs the same way every time.
- The 24 revenue skills split into six stages of four: ICP and sourcing, inbox infrastructure, campaign and copy, LinkedIn and content, pipeline and deals, and retention and expansion.
- The grounded copy skill scrapes each prospect's real website before drafting so the model names actual services rather than inventing plausible ones.
- The stack underneath did not change: Prospeo, Apify, LinkedIn, ScaledMail, Email Bison, Google, Firecrawl, Supabase, HeyReach, Notion, Grain, and Slack.
- Skills rot when vendors change endpoints, and a written bad process runs faster and more consistently than an unwritten one, so client-facing skills carry a human approval gate.
Reviewed and updated August 10, 2026
I Built 99 Claude Code Skills for RevenueFlow. Only 24 Touch Revenue.
The other 75 are admin and engineering. Useful, but they do not book meetings.
I am starting with the ratio because the headline number flatters us, and the interesting part of this is the 24, not the 99. Those 24 are the go-to-market system, and they are the closest thing I have to a written answer for what an AI-operated revenue function actually looks like at the file level.
What a skill is
A skill is a folder with a SKILL.md file in it.
That file is the process, written in plain English: when to use it, what inputs it needs, the exact steps in order, the guardrails, the approval gates, and what finished looks like. You call it like a command and it runs the same way every time.
There is no fine-tuning, no orchestration framework, no agent marketplace. The whole technology is a text file the model reads before it starts. Which is why the unlock is not a smarter model. It is a written process that runs the same way whether I run it or someone who started last week does.
The 24, grouped by go-to-market stage
1. ICP and lead sourcing
/lead-sourcing, /campaign-targeting, /decision-maker-finder, /hiring-signal-leads
Score the vertical before you spend on it, then build the list against it. The ordering is the point. Targeting decisions are cheap to change on a document and expensive to change after you have enriched 4,000 records, which is the argument for settling the targeting criteria before the first pull rather than after it.
2. Inbox infrastructure
/scaledmail-purchase, /warmup-setup, /health-diagnostics, /domain-replacement
Buy pre-warmed domains, set the sending limits, and pull dying domains before they burn the rest of the pool. This is the least interesting group and the one that has saved us the most money, because a domain replaced in week two costs a few dollars and a domain replaced in week eight costs a campaign.
3. Campaign and copy
/email-copy, /campaign-enrichment, /grounded-creative, /email-copy-modifier
The grounded one is the most instructive skill we have. It scrapes each prospect's real website before drafting, so the model names actual services the company sells instead of inventing plausible ones.
That single constraint changed the failure mode of AI copy for us. An ungrounded model writes fluent sentences about a business it has guessed at, and the guess is wrong often enough that a good prospect reads one line and concludes nobody looked. A grounded model has to work from what is on the page, so when it has nothing specific to say, it says something generic instead of something false. Generic is survivable. Wrong is not. The same logic sits behind eight workflows that turn Claude Code into a cold email machine.
4. LinkedIn and content
/heyreach-campaign, /setup-reply-agent, /linkedin-flywheel, /heyreach-analytics
Second channel, same lead data, no manual rebuild. The reason this is four skills rather than one is that LinkedIn has a stricter error cost than email. A client can open the tool and read every message you queued, so the review step is written into the file rather than left to whoever runs it.
5. Pipeline and deals
/deal-flow, /meeting-prep, /icp-qualification, /sales-role-play
Turn a reply into a booked call without guessing who is worth the hour. Reply triage is where most automated outbound quietly stops being automated, because teams automate the sending and leave a person reading a shared inbox twice a day. The system then runs at machine speed right up until the moment a prospect says yes.
6. Retention and expansion
/campaign-analytics, /client-onboarding, /client-rescue-insights, /upgrade-radar
Catch a flat client in week two instead of month two. Nothing here creates pipeline. All of it protects pipeline that already exists, which is why it was the last group to get written and the group I would build first if I were starting again.
Why the written process beats the better model
Every few months a new model arrives and does the same job with fewer mistakes. That is real, and it is not the constraint.
The constraint is that the process for operating the stack used to live in one person's head. When that person was on a call, the work waited. When that person described the process out loud, they left out the exceptions, because the exceptions are the part you stop noticing after the fiftieth time.
Writing it into a file forces the exceptions to the surface. Half of every SKILL.md we have is not the happy path. It is the list of things that go wrong, what to check first, and the point at which the agent must stop and ask a human. That list is the asset. The model is interchangeable.
This is the same shift described in Claude Code as the best GTM tool of 2026: the tools already worked, and the thing that changed is who operates them.
What breaks
Three honest failure modes, all of which we have hit.
Skills rot. A vendor changes an endpoint or a pricing tier and the file is now confidently wrong. A stale skill is worse than no skill, because it runs anyway. We treat any skill that has not been executed in a quarter as unverified.
A bad process runs faster. Writing down a flawed workflow does not fix it. It makes it consistent, repeatable, and much harder to notice, because the output arrives with the same confidence as the good ones. Every skill that touches a client has a human approval gate written into it for exactly this reason.
The 75 do not disappear. Most agent work is still admin. Deploy steps, repo hygiene, infrastructure checks, reporting. That work is worth automating and it does not show up in revenue, so anyone counting skills as a proxy for pipeline is counting the wrong thing.
The stack underneath
Sourcing runs on Prospeo, Apify, and LinkedIn. Sending runs on ScaledMail, Email Bison, and Google. Copy runs on Firecrawl and Supabase. LinkedIn runs on HeyReach and Notion. Deals run on Grain and Slack.
None of that changed this year. Every one of those tools was already in the stack before the skills existed, and we did not replace a single vendor to make this work. If you want the API-level view of the same stack, 22 APIs holding an outbound stack together is the most useful map we have published of it.
How to start
Pick the workflow that wastes the most of someone's week and write it down as you actually do it, exceptions included. Run it read-only first and review the output as a file. Only after the output is right three times in a row should it write to a live system.
Then write the second one. The compounding is real, because every skill after the first reuses the same credentials, the same ICP context, and the same guardrails.
If you want narrower starting points, we have written up five RevOps workflows you can run from the terminal, five marketing workflows, the wider B2B marketing set, and the features most people never find. And if you are wondering who owns this inside a company, that role now has a name, covered in what a GTM engineer does.
Ninety-nine folders. Twenty-four that matter. The count is not the point. The writing down is.
RevenueFlow builds AI-native pipeline systems and you pay per qualified meeting, not a retainer. No paying for activity. You only pay when we book you a qualified sales meeting. See if you qualify.
Frequently asked questions.
Frequently asked questions- What is a Claude Code skill?
- A folder with a SKILL.md file inside it. The file describes a process in plain English: when to use it, what inputs it needs, the steps in order, the guardrails, the approval gates, and what finished looks like. You call it like a command and the agent follows the same path every time, with no fine-tuning or framework involved.
- How many Claude Code skills does a GTM team actually need?
- Far fewer than 99. We run 24 that touch revenue, grouped four each across sourcing, inbox infrastructure, copy, LinkedIn, pipeline, and retention. Most teams get real value from two or three, starting with whichever workflow wastes the most of someone's week, usually list building or reply triage.
- Does using Claude Code for GTM mean replacing your existing tools?
- No. Every tool in our stack predates the skills, and we replaced none of them to make this work. Prospeo, Apify, ScaledMail, Email Bison, Firecrawl, Supabase, HeyReach, Grain, and Slack all kept doing their jobs. What changed is that the process for operating them stopped living in one person's head.
- What goes wrong when you automate GTM workflows with agent skills?
- Three things. Skills go stale when a vendor changes an endpoint, and a stale skill runs anyway with full confidence. A badly designed process becomes consistent and harder to notice once written down. And most agent work turns out to be admin rather than revenue, so counting skills is a poor proxy for pipeline.
- How do you stop AI from inventing details in cold email copy?
- Ground it in the prospect's own website before it drafts. Our copy skill scrapes each company's real pages so the model can only name services that exist. When it has nothing specific to work with, it writes something generic instead of something false, and generic is survivable in a way that a wrong claim is not.
About the author.
Tim Carden is CMO / CTO at RevenueFlow, which builds and operates outbound revenue engines for B2B companies. Studied at McGill University.
Tim Carden · CMO / CTO
Connect on LinkedIn →Explore more.
Ready to scale your outreach?
We build GTM engines that book real meetings. See the receipts.
Related articles.
Three Buyer States, Three Plays: Demand Creation, Capture, and Conversion
Every B2B buyer sits in one of three states: not aware, shortlisting, or ready to buy. Each needs a different play, a different stack, and a different metric.
ZoomInfo Renamed Its Ticker to GTM. Fourteen Months Later It Trades Under $1B.
ZoomInfo took the ticker GTM in May 2025. By late July 2026 it traded under $1B. Here is what the GTM software category is actually worth, with dates attached.