Module 02 -- Build it, run it, refine it

Building with Skills, Agents, and Plugins

Wire 3 features together into one running workflow: a Skill, the right runtime, and a Plugin, anchored in a real PE workflow. About 75 minutes lab.

LabTBC, week of May 18 or 25

What we cover

Build (or import) a Skill for one PE workflow. The lab anchor: a portfolio company news summarizer that takes the last 24 hours of headlines for a named portco and condenses them into a 3-bullet brief.
Connect an Agent that runs the Skill on a recurring trigger. In Cowork, that means a scheduled task only when the computer is awake and Claude Desktop is open. If the job must run while the laptop is closed, use platform automation, Claude Code /schedule, or another non-Cowork runtime. Same news summarizer, different runtime depending on the need.
Install a Plugin from Anthropic's marketplace that adds capability the Skill needs. For the news workflow, a CRM or knowledge base Plugin so the daily output writes to a portco page instead of sitting in a chat window.
Wire the 3 together: input arrives, Skill processes, Agent decides what to do next, Plugin persists. End-to-end on one of your real portcos.
Concrete walkthrough you can copy: paste the exact Skill instruction (one paragraph naming inputs, output format, and constraints), choose the runtime, and name the Plugin target (a CRM record or knowledge base page where the daily brief lands). End-to-end, one portco, run live in the room.
Per-step checkpoints to keep the chain honest. After Skill build: run it once manually and confirm the output shape matches what you described. After scheduled-task setup: dry-run the trigger while Desktop is open, watch what happens. After Plugin install: write one test entry by hand and confirm it lands where you expect. If a checkpoint fails, fix before chaining the next step -- chain-and-pray costs more time than checkpoint-and-fix.
Run it once with you watching. Run it again on a different portco to confirm it generalizes.

Why it matters

A Skill by itself is a slightly faster prompt. A Skill plus the right runtime plus a Plugin is the difference between you reading news every morning and a reviewed brief landing where the team can use it. The leverage is in the combination, not any one feature.

Hands-on moment

Pick one of your portcos. Build the daily news summarizer end-to-end during the lab. If using Cowork scheduled tasks, set it to fire shortly while Claude Desktop is open so you watch the first output land before you leave the room.

Peer moment (3 min)

“Compare the output your Agent produced with a neighbor's. Whose summarized better, and what was different about the Skill instruction that explains the gap?”

Workbook steps

Build (or import) a Skill for one PE workflow. The lab anchor: a portfolio company news summarizer that takes the last 24 hours of headlines for a named portco and condenses them into a 3-bullet brief.

Connect an Agent that runs the Skill on a recurring trigger. In Cowork, that means a scheduled task only when the computer is awake and Claude Desktop is open. If the job must run while the laptop is closed, use platform automation, Claude Code /schedule, or another non-Cowork runtime. Same news summarizer, different runtime depending on the need.

Install a Plugin from Anthropic's marketplace that adds capability the Skill needs. For the news workflow, a CRM or knowledge base Plugin so the daily output writes to a portco page instead of sitting in a chat window.

Wire the 3 together: input arrives, Skill processes, Agent decides what to do next, Plugin persists. End-to-end on one of your real portcos.

Concrete walkthrough you can copy: paste the exact Skill instruction (one paragraph naming inputs, output format, and constraints), choose the runtime, and name the Plugin target (a CRM record or knowledge base page where the daily brief lands). End-to-end, one portco, run live in the room.

Per-step checkpoints to keep the chain honest. After Skill build: run it once manually and confirm the output shape matches what you described. After scheduled-task setup: dry-run the trigger while Desktop is open, watch what happens. After Plugin install: write one test entry by hand and confirm it lands where you expect. If a checkpoint fails, fix before chaining the next step -- chain-and-pray costs more time than checkpoint-and-fix.

Run it once with you watching. Run it again on a different portco to confirm it generalizes.