← sources

Every day there’s a new skill or tool. At what point does it become too much?

source: reddit (reddit:ClaudeCode)published 2026-09-27 18:26ingested 2026-09-28 02:41

retained text

I keep seeing more stuff around skills lately, especially meta-skills that create/improve other skills or people having Claude turn repeated workflows into reusable skills.

And honestly I like the idea a lot. I also started doing the same thing.

But then I had kind of the opposite problem.

The better this works, the more skills you end up with.

I saw someone here yesterday saying they have 77 skills already, another person has a system that identifies when Claude is repeating work and creates a skill for it, people are keeping feedback inside skills so they improve after mistakes etc.

So let's say this actually becomes normal.

You have 20 skills now. Then 100. Maybe eventually hundreds. Add MCP servers, plugins, agents and tools on top.

How is the agent supposed to know what it should actually use for one task?

Claude already does progressive loading for skills, which is obviously much better than loading every full SKILL.md. But it still needs the names/descriptions of installed skills to decide what is relevant, and MCP tool schemas have their own context cost too.

I ran into this problem with my own setup because I use more than one coding agent and had skills/tools spread around different places.

So I started building something called Lockkeeper.

The basic idea is pretty simple: keep the big capability library outside the agent, then route each task to a small set of skills/MCPs/tools that actually look relevant.

So instead of the agent having to choose from everything you installed, it gets something like 5-10 capabilities for this specific task.

I wanted to see if this could work without putting another LLM in front of the LLM, so I tested the router against SkillRouter Eval Core.

With 26k real skills in the pool, the current local ranker puts a correct skill at #1 around 65% of the time.

With the full \~79k pool it's around 55%.

No model/API needed for the default routing.

Those numbers are not some "routing is solved" claim btw. It's only 75 tasks and I developed some of the ranking changes against the same benchmark, so it's not a proper held-out result. I put the methodology and scripts in the repo so it can be reproduced.

The other thing that started bothering me is security.

Once we get to the point where people have hundreds of skills from GitHub/marketplaces and agents automatically decide which instructions to load, [SKILL.md](http://SKILL.md) is basically untrusted instructions going straight into your agent.

So I added an audit layer for prompt injection, credential exfiltration patterns, hidden payloads etc. Still something I want people to attack/break because static scanning definitely isn't magic.

Right now Lockkeeper works with Claude Code automatically through a hook, and as an MCP server for Codex/Cursor/etc.

Repo: [https://github.com/Hannay001/lockkeeper](https://github.com/Hannay001/lockkeeper)

It's pretty early. I have basically the opposite problem from a mature project right now: it works, but I don't have enough people with horrible real-world setups using it 😅

So if you have 30/50/100+ skills, lots of MCPs, or you use multiple coding agents, I'm actually interested in how you're dealing with this.

Do you keep everything installed and let Claude choose? Keep a tiny curated list? Have your own routing system? Or does this only become a problem once you go completely overboard?

document fetched 2026-09-28 00:37 · canonical https://reddit.com/r/ClaudeCode/comments/1wrrlf0/every_day_theres_a_new_skill_or_tool_at_what