Scott ran two Claude Code skills live. One turns a call transcript into structured buyer insight. The other finds accounts, grabs contacts, verifies emails, and pushes a campaign to Instantly. All from natural language. Zero UI clicks.
▶️ Watch the Full Recording
Guest Speaker
Founder at Driver ROI • Claude Code Coach at Clay Bootcamp
Scott came up through finance and marketing tech, and calls himself "semi-technical" (intermediate SQL, not a Python or JavaScript coder). He started using Claude Code as his way into building GTM tools without having to learn a new stack, and he's been shipping reusable skills ever since. His angle: orchestrate the tools you already use, don't replace them.
Outside of Clay Bootcamp, Scott runs Driver ROI. The two skill files he walked through in this session are posted below so you can copy, edit, and run them yourself.
Session Overview
Drop a call transcript into the terminal and a single skill returns pain points with direct quotes, buying triggers, objections, competitive intel, decision criteria, and a drafted follow-up email.
One natural-language brief. Claude finds ICP accounts through DiscoLike, pulls decision-maker contacts, verifies the emails, writes sequence copy with personalized variables, and stages the campaign in Instantly ready to launch.
Every workflow writes into a GitHub-backed repo of tagged markdown files. Positioning, objections, competitive intel. Claude traverses that context on every future run so you build on what you already know.
Scott's framework for getting here. Start manual. Turn the manual flow into a skill. Put the skill behind a cron. Eventually let Claude iterate on itself. One step at a time, not all at once.
The Skills
Each one is a single Claude Code skill file. The first turns raw call transcripts into structured buyer intel. The second runs a full ICP-to-campaign prospecting flow without a UI. Both files are posted on Scott's site, ready to copy and edit.
Drag a call transcript into the terminal, invoke the skill, and get back participants and roles, pain points with direct quotes, buying triggers, objections, competitive intel, tech stack, decision criteria and timeline, plus a drafted follow-up email. The skill file is editable, so you can swap in any discovery framework you already use (MEDDIC, SPICED, whatever). Scott's next move is to wire this to a cron that pulls last week's Gong or Fathom transcripts automatically every Monday at 8:30am.
Natural-language brief: "25 accounts in construction SaaS this size in the US, and create a campaign with 3 steps and custom variables." The skill hits DiscoLike via MCP to find accounts, pulls VP+ revenue-leader contacts, verifies the emails in a second pass (DiscoLike's native emails bounce higher than Scott wanted), writes subject lines and personalized body copy, sets campaign defaults (send volume, mailboxes), and stages the campaign in Instantly ready to launch.
These are the exact markdown files Scott ran in the session. Open, copy the contents into your own project's .claude/skills/ folder, edit, and run.
On Scott's build list
Two things Scott has queued up, both building on the skills above. Useful pattern references if you want to take your own workflow past "manual skill" to "the system runs itself."
A cron hits every Monday at 8:30am, pulls last week's CRM transcripts, runs /extract-insights, and updates the positioning, objections, and competitor docs in the Context OS repo. No human in the loop; the doc set stays current.
250 test accounts split A/B. A gets Scott's copy. B gets Claude's proposed variant. After 48 hours Claude reads the delta on interested-reply rate and iterates. Inspired by karpathy/autoresearch, adapted to outbound copy. The loop closes itself.
Principles
Don't try to build the async self-learning end state in one go. Claude will struggle and so will you. Start manual. Turn the manual flow into a skill. Put the skill behind a cron. Each step earns the next one.
The moment Scott realized he was telling Claude "move this button here, make sure that button does that," he dropped the UI entirely. Claude Code orchestrates the tools you already pay for. The UI is for humans who need it, not for machines that don't.
Pick one workflow. Ship it end to end. Then pick the next. Trying to build a dozen skills in parallel gives you a dozen half-skills that don't compose. The Context OS repo grows tighter as you focus.
Every skill writes its output to a GitHub-backed repo of tagged markdown files. Call notes, objections, positioning, competitive intel. Claude traverses that graph on every run, so future work builds on what you already figured out.
"I'm not going to vibe-code a CRM." Keep Claude Code on orchestration. Don't rebuild the surfaces people already live in (HubSpot, Salesforce, Gong). Wire into them; don't replace them.
Scott's one bad incident deleted about 200 contacts from a campaign in YOLO mode. The fix is a hard rule: any destructive action has to ask first. Skip that rule once and you'll buy the lesson the same way he did.
Q&A Highlights
Stack
Resources
The next Claude Code cohort runs for one week. You build real workflows, get real feedback, and walk out with skills you can run the next day.
Join the Cohort →Clay Bootcamp • claudecode.claybootcamp.com
Stay in touch
Reach out to either of us on LinkedIn. Always happy to chat.
