I was offered 292 AI skills. I installed 21.
Halfway through building ParkEase, I found a popular open-source pack for Claude Code: 292 skills, 68 specialist agents, 94 slash commands and a hook runtime. One install command and my AI agent would know about everythi
Halfway through building ParkEase, I found a popular open-source pack for Claude Code: 292 skills, 68 specialist agents, 94 slash commands and a hook runtime. One install command and my AI agent would know about everything.
I installed 21 of the skills and 25 of the agents. Here's why the other 271 skills were a liability, not a bonus.
More skills, worse choices
An agent picks skills by reading their short descriptions and matching them to the task. Every installed skill is a candidate at every decision.
The full pack would have taken my skill index from 56 to 348, including skills for Django, Laravel, Rails, Spring Boot, .NET, Flutter, Vue, Angular, MySQL and Prisma. ParkEase uses none of those. Each one is a plausible-looking wrong turn: a confident, well-written instruction file for somebody else's stack.
Selection noise isn't harmless. It's how an agent ends up following excellent advice written for a different project.
The most dangerous skill was the most relevant-looking one
The pack had a skill for NestJS, which is exactly the framework ParkEase's API uses. It looked like the obvious one to install.
It contradicted four of my architecture decisions at once:
| The skill says | ParkEase says |
|---|---|
Organise by src/modules/
|
Role folders for controllers, domain folders for logic |
Put shared guards in common/
|
There is no common/ folder |
| Validate DTOs with class-validator | One Zod schema, shared by four consumers |
| Isolate transactions "for Prisma or TypeORM" | Drizzle ORM |
Install that and every API task becomes an argument between my rules and a skill that sounds authoritative. Worse, a newer session might not argue at all.
So it went on a written quarantine list, along with skills that duplicated gates I already had, and two whose names collided with skills I already used. One of those, called taste, turned out to be about music-video creative direction. Its name alone would have pulled it into UI work.
How I curated
- Vetted every candidate by hand against the stack and the decision records.
- Installed project-local, inside the repo, not globally, so nothing leaks into my other projects.
- Verified by count after installing. The installer pulled in transitive dependencies, so the process was install → prune → re-audit until the numbers matched what I'd decided.
- Pinned the hook profile to minimal. The pack's default hook profile blocked ordinary shell commands until I changed it.
- Wrote it all down in a decision record, so a future session wouldn't "helpfully" install the rest.
The rule I'd give anyone
Skills are context. Context is attention. Install the ones whose advice you'd follow, for the stack you actually have, and treat everything else as noise with good grammar.
Next: the most expensive mistake of the build. One task, about forty AI reviewer agents, and most of a week's usage limit gone.
ParkEase is a peer-to-peer parking marketplace for India, built solo with Claude Code.
Code: https://github.com/Deonkar/parkease
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.