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

Building Skills Safely

Build a Skill yourself, end-to-end, with Claude's Skill Builder. Then the policy: why GGP standardizes on Anthropic's marketplace plus the firm's admin-curated private one. About an hour lab.

LabTBC, week of May 18

What we cover

Pick a real PE workflow you would like to automate -- for this lab, a confidential information memorandum (CIM) section extractor (pull financials, management, and market sections from a CIM PDF into a structured table).
Open Claude's Skill Builder in Cowork and describe the workflow in plain language: what goes in, what comes out, what the in-between steps are.
Test the Skill on the public S-1 filing linked in the deck. Read the output the way an analyst reads draft work.
Iterate the Skill description based on what the output got wrong. One round of edits beats 3 starts from scratch.
Before/after prompt pair, worked side-by-side: a CIM extractor with no guardrails (output drifts -- pulls the wrong table, invents columns, missing rows) versus the same Skill with named columns, output format spec, and a 'do-not-extract' list (output holds). The gap between the two is where the Skill description earns its keep.
Three common failure modes for analyst-built Skills, with a fix pattern for each: description drift (the Skill mutates from one run to the next -- fix by pinning the description verbatim before testing); hallucinated structure (the model invents fields not in the source CIM -- fix by adding a 'use only these column names' constraint); silent partial outputs (the Skill 'succeeds' but missed half the data -- fix by adding a row-count sanity check at the end). Each is a re-test moment, not a re-build moment.
Save the Skill locally to your Cowork instance and confirm it works as a /command for you. To ship it firm-wide on Cowork Enterprise: submit the Skill to the GGP plugin owner; the admin reviews the Skill code, packages it inside a Plugin, and pushes the Plugin to the firm's private marketplace. There is no in-product self-publish workflow on Enterprise today -- the admin gate is the security review.
Before you build from scratch, check Anthropic's pre-built vertical plugin packs (Private Equity, Investment Banking, Financial Analysis, Equity Research, Wealth Management, announced 2026-02-24). One of those packs may already cover the workflow.
The policy at the end of the lab: Skills can include arbitrary code. Downloading from Reddit, random GitHub repos, or community forums is a malware vector. Anthropic's marketplace vets contributors. GGP standardizes on Anthropic's marketplace plus the firm's own admin-curated private marketplace -- nothing else.

Why it matters

Skills are the unit of compounding firm capability. The same Skill an analyst builds on a Tuesday should run for every analyst on Wednesday. That only works if the Skills you ship come from a place you trust.

Hands-on moment

Build the CIM extractor live. Run it on the public S-1 filing. Save it locally. By the close of the lab, you have shipped your first Skill -- ready to submit to the firm's plugin owner for review and firm-wide distribution.

Peer moment (3 min)

“Trade the output your CIM extractor produced with a neighbor. Did their Skill catch a section yours missed? Borrow the better instruction and re-run.”

Workbook steps

Pick a real PE workflow you would like to automate -- for this lab, a confidential information memorandum (CIM) section extractor (pull financials, management, and market sections from a CIM PDF into a structured table).

Open Claude's Skill Builder in Cowork and describe the workflow in plain language: what goes in, what comes out, what the in-between steps are.

Test the Skill on the public S-1 filing linked in the deck. Read the output the way an analyst reads draft work.

Iterate the Skill description based on what the output got wrong. One round of edits beats 3 starts from scratch.

Before/after prompt pair, worked side-by-side: a CIM extractor with no guardrails (output drifts -- pulls the wrong table, invents columns, missing rows) versus the same Skill with named columns, output format spec, and a 'do-not-extract' list (output holds). The gap between the two is where the Skill description earns its keep.

Three common failure modes for analyst-built Skills, with a fix pattern for each: description drift (the Skill mutates from one run to the next -- fix by pinning the description verbatim before testing); hallucinated structure (the model invents fields not in the source CIM -- fix by adding a 'use only these column names' constraint); silent partial outputs (the Skill 'succeeds' but missed half the data -- fix by adding a row-count sanity check at the end). Each is a re-test moment, not a re-build moment.

Save the Skill locally to your Cowork instance and confirm it works as a /command for you. To ship it firm-wide on Cowork Enterprise: submit the Skill to the GGP plugin owner; the admin reviews the Skill code, packages it inside a Plugin, and pushes the Plugin to the firm's private marketplace. There is no in-product self-publish workflow on Enterprise today -- the admin gate is the security review.

Before you build from scratch, check Anthropic's pre-built vertical plugin packs (Private Equity, Investment Banking, Financial Analysis, Equity Research, Wealth Management, announced 2026-02-24). One of those packs may already cover the workflow.

The policy at the end of the lab: Skills can include arbitrary code. Downloading from Reddit, random GitHub repos, or community forums is a malware vector. Anthropic's marketplace vets contributors. GGP standardizes on Anthropic's marketplace plus the firm's own admin-curated private marketplace -- nothing else.

Workbook bundle

Print or open these alongside the session. They are the working surface for the steps above.

labSkill Builder lab worksheet

Related artifacts

labSkill Builder lab worksheet../../course-content/module-2-building-skills-safely/lab-worksheet.md
pre-readAnthropic Skills Authoring Playbook (32 pages, PDF)https://resources.anthropic.com/hubfs/The-Complete-Guide-to-Building-Skill-for-Claude.pdf

Source files live alongside this site under clients/greenridge-growth/; paths above are relative to the syllabus-site root.