The Prompt and Skill Libraries I've Built Since 2025
Eight libraries of prompts and Claude skills, from one long document of prompts to a plugin a team could install, a company marketplace and a public one. An overview of what each held and how I kept them in order.
- My role
- Designed each library and its rules, built or adapted most of the skills, and ran the company marketplace.
- Team
- Mostly me with Claude Code. In the company marketplace, many skills were written by colleagues.
- Timeline
- 2025 to Nov 2026
- Tools
- Claude Code, Claude plugins and marketplaces, Codex, Python, GitHub Actions
The problem
Every time I started work with AI, I explained the same things again. How to research a company before a call. How I write a spec. Which files never to touch. At first I saved those explanations as prompts in a document and pasted them in. Later Claude got skills: a folder of saved instructions that Claude loads on its own when the work calls for it. A plugin bundles skills so someone else can install them in one step.
I started with prompts for my own work at Cadre, the AI consultancy I co-founded. Then I packaged them for our team, then for the whole company and the organizations we trained, then in public. This page is the short version of each.
The libraries, in order
A prompt library in a document (2025). About a dozen long prompts for consulting work, kept in one Word document. Each one ran a whole task: research a company before a first call, pull the decisions out of a meeting transcript, check a client's claims against outside sources, or turn an assessment into a pricing sheet. Several produced a set of finished documents at once, not a single answer.
My own setup (from July 2025, private). My Claude Code configuration, kept in git. It grew into the working rules I give every session, safety hooks that block destructive commands, and a chain of skills for consulting work, from discovery through delivery. In February 2026 I had parallel Claude agents audit it. One finding: the rules that decided when a skill switched on fired wrongly about half the time, so I rebuilt them.
A plugin for my team (November 2025). My setup, packaged so colleagues could install it. It had hooks that stop risky commands and protect files holding passwords and keys, ten commands that take work from plan to review to ship, seven helper agents and ten skills. The getting-started guide was written for people who don't code.
A skills gallery (January 2026). A searchable website of about 120 skills, with filters, side-by-side comparison and one-click downloads. Most of those skills were collected from others, including Anthropic and open-source authors. My part was the curation, the skills written for our own work, and the site.
Folio (February to March 2026). A web app where people could find a prompt or skill, fill in the blanks and copy it, without installing anything. It went live in March with 112 prompts and later added 143 skills, but it never launched to the team. The Folio page tells that story.
Prompt libraries for organizations we trained (2026). When we trained a team in Claude, a prompt library was one of the things they could keep, alongside a Claude coach. By June 2026 I'd rebuilt the library from small pieces you stack: one sets the role, one sets how Claude works the problem, one shapes what comes back, and an optional one raises the bar. About twenty pieces made up a shared core, and an organization could get a layer for its own work on top. The edition for a law firm added pieces for its legal practice and its rule that client details are anonymized before they reach Claude.
The company marketplace (June to August 2026, private). Our team's official place to get skills. It installed into Claude Code, the Claude app and Cowork. It held six plugins grouped by kind of work, such as client work and training. Each plugin had a "lab" twin where new skills waited until they'd been used for real. It held between 34 and 46 skills over the summer, and colleagues wrote many of them. Before the first 36 skills shipped, a scanner and a set of Claude agents removed every real client and colleague name from them.
bs-skills (November 2026, public). Twelve skills in four plugins, for jobs like writing specs and product plans, handover notes between AI sessions, tightening drafts, interactive explainers, and one for songwriting. They work in Claude and in Codex. A few personal skills, like the one that writes in my voice, stay in a separate private library.
How I keep a library healthy
- Each skill lives in one place. If a second version is needed, it gets a new name on purpose.
- Every skill has a change log and a version number, so you can see what changed and when.
- A check blocks any commit that names a client or includes a path from my computer.
- New skills start in a lab. They move into the main plugin after real use, and old ones are retired, not deleted.
- Rules come from evidence. Our marketplace had a rule that skill descriptions must be under 200 characters. A survey of 841 of Anthropic's own skills showed that rule was wrong, so we dropped it.
What I didn't measure
I have no install or usage numbers for any of these. The automated checks confirm a skill is well formed. They don't test whether it switches on at the right moment. That's the first thing I'd add next time.
Try them
The public skills are at github.com/benshapyro/bs-skills. In the Claude app, open Customize, then Plugins, then Add marketplace, and enter benshapyro/bs-skills. In Claude Code:
claude plugin marketplace add benshapyro/bs-skills
claude plugin install product@bs-skills