Source: raw/Claude_Cowork_Just_Changed_Sales_Forever.md; raw/Build_Agent_Teams_within_Claude_Cowork_in_17_min.md; raw/CLAUDE_COWORK_FULL_COURSE_2+_Hours.md (agents and sales chapters, which repeat the two videos above)

Creator: Ben AI · URL: https://www.youtube.com/watch?v=EqJoui72QrU · Duration: 15:08 · Platform: YouTube · Uploaded: 2026-03-19 Also: Build Agent Teams within Claude Cowork in 17 min (https://www.youtube.com/watch?v=A6RbawFHC80, 17:37, uploaded 2026-03-04) · CLAUDE COWORK FULL COURSE 2+ Hours (https://www.youtube.com/watch?v=2HyBA-wkWsA, 2:28:52, uploaded 2026-04-04)

Ben AI’s sales team runs prospecting, lost-lead follow-up, call prep and sales analytics as Claude Cowork skills, each built by doing the work once with Claude and saving it. For lists of 100-200 leads he has Cowork split the work across parallel sub-agents, then packages the qualification and enrichment steps as a plugin with agent instruction files and one command that runs them in sequence. The “agent teams” video title is loose: what it shows is Cowork’s isolated sub-agents, which Ben contrasts with Claude Code’s agent teams. Recorded in March and April 2026, before the chat/Cowork merge; the scheduling limits he describes have since changed.

Key Takeaways

  • Start from Anthropic’s sales plugin, then customise or replace it. Install it under Browse plugins; it ships account research, call prep, call summary and competitive-intelligence skills. Ben calls them generic. To adapt one: “I want to customize the Anthropic call prep skill according to how my company does call preps,” then give the context and what to change. His team’s own sales plugin has skills for CRM prospect mining, email personalisation, lead intelligence, lead qualification, LinkedIn post engagers, pipeline review, sales-rep analysis and win-loss analysis.
  • The connector stack he recommends for sales: Apify (scrapers for LinkedIn profiles, posts and post engagers, X, Instagram, and some lead databases such as Apollo), a lead database (Apollo, Clay, or Vibe Prospecting, which he describes as an API built for AI agents), an outreach tool (Instantly for email; Unipile for LinkedIn, WhatsApp, Facebook and X), the CRM (his is Attio), a meeting transcriber (Fireflies), calendar and email.
  • Build each skill from one manual run. Corrections you make during the run end up in the skill. In his LinkedIn-engager build Claude first qualified leads without scraping their profiles; he told it to use an Apify profile scraper for the About section, company name and website, and the saved skill kept that step.
  • Say “parallel sub-agents” and give batch sizes. Ben says that without the explicit instruction Cowork sometimes runs sub-agents one after another, or uses none. He finds 5-15 items per sub-agent works best; more per agent is slow, and too many agents “runs into complications.”
  • Sub-agents keep the main context clean. Each one researches its batch with its own tools and returns only a summary, so the main agent is not filled with raw research. Ben says this made bulk work on 100-200 leads practical in Cowork, where a single chat could not do it before.
  • Limits he names: sub-agents use a lot of tokens (he suggests the Max plan for regular use), and above roughly 100-200 items the runs get long, complicated and expensive. For large non-human-in-the-loop bulk jobs he still prefers n8n or Make.
  • Cowork sub-agents versus Claude Code agent teams. In Cowork the sub-agents are isolated and report only to the main agent; Claude Code’s agent teams can message each other and share a goal or to-do list. Ben treats agent teams as mainly an engineering feature and says isolated sub-agents suit most sales, marketing and support work and use fewer tokens. See Agent Teams and Subagents.
  • Schedule the daily ones. Call prep each morning and pipeline review each day are his candidates; win-loss analysis monthly.

Five sales skills he shows

1. LinkedIn post engagers (prospecting). Invoked as /linkedin-post-engagers with a post URL and qualification criteria. On Chris Long’s latest post it found 127 unique engagers, qualified 17, enriched them with company size and description, and returned a CSV with LinkedIn URL, About text, company size, website, description and the post they engaged with. The build prompt: “scrape the most recent 10 posts of [profile], extract the post commenters and engagers data, and present them in a CSV along with the LinkedIn profile URL, name, headline, and the post that they engaged with” (he named the Apify actor to save time; he says Claude can find one itself). The saved SKILL.md: scrape posts with engagement data (named Apify actor) → extract and deduplicate engagers → ICP qualification → full LinkedIn profile scrape → company LinkedIn page scrape → CSV. Outreach is a separate skill.

2. CRM prospect miner (lost-lead follow-up). Reads the Lost column of the CRM, asks for opportunity criteria, finds each lead’s LinkedIn profile and scrapes it with Apify, reads past email threads, and returns a CSV ranking lost leads by company size and ICP fit with a summary of past communication. On his CRM it found 160 records. The build instruction included “launch sub-agents for this and give each sub-agent 15 leads to analyze.” Connectors: Gmail, Apify, CRM. Drafting and sending follow-ups through Gmail is the next step.

3. Call prep (customised Anthropic skill). Finds the lead in the CRM, reads email history and past Fireflies transcripts, researches the company on the web, and produces an HTML brief: company overview, account snapshot, interaction history, suggested agenda, discovery questions. Scheduled version: “every morning at 7:00 a.m., check my CRM and go through all of the meetings I have today and use the call prep skill for each of the meetings I have today, present all the call prep briefs in a nice HTML dashboard.”

4. Win-loss analysis and rep performance. Win-loss reads the Won and Lost columns, the meeting transcripts and the emails for every lead, and writes a Google Doc: win rate, best and worst customer profiles, common objections, behaviours that predict outcomes (his report found engagement depth did), deals by company type and price, case studies of won deals, key differentiators, red flags and early disqualification criteria. Rep analysis grades one rep from transcripts, CRM data and emails: overall grade, top three strengths and improvement areas, a per-client scorecard, grades by stage (discovery, demo, objection handling, pricing, follow-up), and five recommendations.

5. Pipeline review (customised Anthropic skill). Asks which pipeline, which stage and where to send the summary; he chose a Slack DM. It lists the leads to work today because they are hot or at risk. He suggests extending it to pipeline hygiene, updating CRM records from emails and transcripts.

Bulk qualification and enrichment with sub-agents

Ben’s demo list: 150 marketing agencies from Apollo. Qualification against “marketing agencies who provide SEO services and are based in the US” ran 15 parallel sub-agents, each using web search, and took about two minutes by his account; 82 qualified. Enrichment ran 18 more (17 lead researchers and one LinkedIn scraper); a personalisation pass ran 17 more and wrote 37 icebreakers.

The prompts, in order:

  1. Qualify. “I want to qualify these leads against my ICP criteria. My ICP criteria are [criteria]. Please spin up 15 parallel sub-agents that each research 10 leads in this list and make sure they’re qualified leads. Based on the summaries you get back, you deliver an updated CSV sheet and add a column with the qualification status.” Output gained qualification-status and qualification-reason columns.
  2. Enrich. “Please spin up 10 parallel sub-agents that enrich seven qualified leads each. They should find the lead’s last LinkedIn post, the company size and the company description using the Apify scraper, specifically the LinkedIn scrapers… add these data points to the CSV sheet so we can later use that data for personalized outreach.”
  3. Package as a plugin. “Please create a new plugin that’s called Ben’s outbound sales process. Inside of that plugin, I want to create one skill for the exact way we did the lead qualification and call it Ben’s lead qualification skill. Make sure to also create an agent.md, and the skill should refer to the agent.md instructions on what the agents need to do. Also, create a skill for the exact way we did the enrichment process… also make sure to create a separate agent.md for the enrichment process. In both of these skills, it should clearly be instructed to spin up parallel batches of sub-agents.”
  4. Add a command. “Also, please create a command.md where I can trigger both of these skills in sequence when I trigger a slash command called outbound enrichment. This command should ask the user for a CSV sheet with leads, and it should then automatically use the lead qualification skill and the lead enrichment skill in sequence to deliver a fully enriched CSV to the user at the end. All of these new documents and files should be stored in the same plugin.”

The result: a plugin with two skills, two agent files and one command. His production version, “outbound pipeline”, adds a personalised-outreach skill and runs all three. If the Install button does not appear or fails, ask for the plugin as a zip and upload it under Plugins → + → Upload plugin; pushing it to a GitHub repo gives a versioned link to share. A command can itself be scheduled.

Outdated details

  • Ben’s scheduled tasks run only while the desktop app is open (“as soon as you open your Cowork in the morning, it will start running”). Since July 2026 scheduled tasks inside cloud projects run with no device online. See Claude Cowork.
  • He describes Cowork as the second tab of the desktop app on paid plans; chat and Cowork merged in September 2026.
  • Whether Cowork sub-agents can now coordinate with each other is not covered by any later source in the wiki; check before relying on the isolation point.

Try It

  1. Install Anthropic’s sales plugin and customise its call-prep skill with your own brief format; schedule it for each morning.
  2. Connect Apify, your CRM, email and meeting transcriber before building anything else.
  3. Run one prospecting workflow by hand on a small list, correcting Claude as you go, then save it as a skill.
  4. For a 100-200 lead list, prompt for explicit parallel sub-agents with 5-15 leads each and a CSV column for the result and the reason.
  5. Package qualification and enrichment as a plugin with agent.md files and one command; export the zip for teammates.
  6. Run a win-loss analysis once on last quarter’s closed deals before scheduling it monthly.

Open Questions

  • Token cost of the 150-lead run is not given; Ben only says sub-agents use a lot and recommends Max.
  • The 100-200 item ceiling and 5-15 per sub-agent are his experience, not documented limits.
  • Accuracy of the qualification. He shows counts (82 of 150; 17 of 127) but not a spot-check of how many qualified leads were correct.
  • Commands inside plugins after later Claude changes. The videos show command.md files; whether plugin commands still work the same way in the merged product is not covered here.