A Prompt Library I Shipped but Never Launched
One place for our team to find a prompt or skill, fill in the blanks and copy it. I built it with Claude Code in a month. Nobody used it.
- My role
- Designed it, wrote every plan and directed the build. All 124 commits are mine, written with Claude Code.
- Team
- Just me and coding agents.
- Timeline
- Feb to Aug 2026
- Tools
- Claude Code, Next.js, Supabase, TypeScript, Vitest, Playwright, Vercel
The problem
At Cadre, the good prompts lived in people's heads, Slack threads and shared docs. When someone needed one, they asked a colleague first and searched Slack second. Everyone knew the prompts existed. Nobody could find them.
So I built Folio: one place to find a prompt, fill in the blanks and copy it into Claude or ChatGPT. Later it held skills too, the saved instructions Claude loads when the work calls for them. The plan was for our team to use it, and for each organization we worked with to get its own private space.
What it did
- Find it fast. One library for prompts and skills, with search, filters and a keyboard shortcut that opens a search box from anywhere.
- Fill in the blanks. A prompt could have blanks like
{{company_name}}. You filled them once, Folio remembered your answers, and the copy button gave you the finished prompt. - A private space for each organization. One organization could never see another's prompts. The database enforced that, not just the app.
- Improve it together. Anyone could copy a prompt and change it, or suggest an edit to the original. The owner could accept part of a suggestion. Every change was kept in a version history.
How I built it
I didn't write the code by hand. I wrote the plan for each feature, and Claude Code built it. There were 82 written plans in total. After every session the agent wrote a handover note so the next one could pick up without me explaining again. That came to 110 notes.
It went live on March 12, 2026, with 112 prompts. Two weeks later I added 143 skills. Then I had agents review the whole codebase for security problems. They found a shortcut in the code that could have skipped the login check, and we closed it. At the end of that review, 370 tests passed.
What happened
Nobody used it. I told a few people about it, but I never announced it, showed anyone how to use it or followed up. The production review ended with "Next: get real users on it." Instead, in June I came back and built more features.
In August I sat down to relaunch it as the place our team would get its skills. By then the team's skills lived in a company marketplace, which my skills and plugins page covers. This time I started with why people weren't using the skills we already had. There were four reasons:
- Finding them. People didn't know what existed.
- Setting them up. Installing a skill was too technical for many people on the team.
- Knowing which one. People didn't know which skill fit the job, or how to use it.
- Fit. Some skills didn't match the work people actually did.
A better library helps with the first three. It can't fix the fourth. That takes better skills. I wrote the relaunch plan, but I left Cadre before it was built.
What I learned
Building it was the easy part. With Claude Code I had a working, tested product in a month. Getting people to change how they work was the real job, and I skipped it. Next time I'd plan the launch before writing the first line: who uses it first, how they hear about it, and what I'd measure in the first two weeks.