Agent Skills: The Ones Worth Keeping After Anthropic Said to Delete Them

Two things happened in the same thirty days, and nobody put them next to each other.

The install genre kept growing: 22 free skills, 14 for Hermes, 10 more, 1000 you can drop into anything, 100 from Google. And in the middle of it, the creator of Claude Code sat down on the Y Combinator channel and told people to delete theirs. The video breaking that down pulled 94,561 views, more than any install list published in the same window.

I run agent skills in two places: a Claude Code setup that maintains this site, and a Hermes gateway that has been up since April. So I went and checked which advice actually holds.

What the creator of Claude Code told you to do with your skills

He told you to delete them, on a schedule, and he meant it literally.

Boris Cherny's line, from the interview: "For people that aren't building agentic products, but you're using Claude Code, every 6 months delete your CLAUDE.md. Delete your skills. Delete your hooks. See what the model does and it might surprise you."

The reasoning is not that skills are bad. It is that most of them are scar tissue. Anthropic deleted over 80% of Claude Code's own system prompt, and Cherny explains why: "A lot of the stuff in the system prompt was correcting for behaviors that the model should have known, but it didn't. Now, Opus 5 just does it." Your skills folder is full of the same corrections, written for a model that no longer needs them.

The corollary he gives is sharper than the delete advice: "a really common mistake I see is people give it way overly specific instructions... You want to describe the task, the guardrails, the exit criteria, and then just let the model cook." And the part almost nobody quotes: "The skill nowadays is less about prompt engineering and more about how do you make it possible for Claude to verify its work along the way. And the verification is probably the single most important thing that people do not get right."

That is the test. Not "is this skill good?" but "is this skill still doing work the model cannot do for itself?"

The agent skills that survive the delete test

Three categories survive, and one does not. Firecrawl's skills roundup names the first two distinctions well (capability uplift versus encoded preference), and the delete test adds the third.

Capability, not instruction. A skill that gives the agent something it genuinely does not have survives every model upgrade, because no model update grants it web scraping, or a document format, or access to your issue tracker. Delete advice does not touch these.

Taste and context only you have. This is the exception Nate Herk found when he actually ran the experiment instead of just repeating the advice. He duplicated his repo, stripped CLAUDE.md and every skill, and re-ran a real task. The stripped version produced better structure: it organised the output into ideas with timestamps, which he preferred. The version with skills produced better branding: his header image, his colours, his links. His conclusion is the usable one: delete the skills that dictate how; keep the ones carrying context the model cannot infer: where your files live, what your brand looks like, what "done" means here.

Verification, because Cherny said it is what people get wrong. The best-documented example is Cole Medin's library, which splits work into an outer loop (PRD, architecture spec, then slicing an epic into tickets) and an inner loop (prime the codebase, plan, implement, validate). His framing: "The most important part of the entire plan document is the testing strategy. I call it test-driven development but for agents. Before we write a single line of code, we plan how to test that code. What we get back from the agent is never its first pass." That is Cherny's exit criteria, written down as a reusable file.

What does not survive: the skill that explains a task the model already understands. Most 22-item install lists are full of these. They were useful against a weaker model and are now overhead in your context window.

Three skills that pay for themselves in tokens

These are the clearest case of capability uplift, because none of them is an instruction: each changes what reaches the context window. Sharbel A. walks through all three in a video with 64,908 views, and his numbers are worth repeating because he separates his own from the repo's.

Defuddle strips a web page to reader-mode markdown before the agent ever reads it. No navigation, no cookie banner, no newsletter popup sitting in your conversation for every turn after.

Caveman compresses output. The repo claims a 65% cut in output tokens, and his explanation of why that compounds is the useful part: "output becomes part of your conversation and gets sent back with every turn after that. Basically, pay for the same question twice, three times, four times."

Codebase memory indexes a repository into a graph so the agent queries a map instead of re-reading files. The project claims 99% fewer tokens for exploration; he ran it himself and got 95%, and says so plainly. That honesty is the reason I am citing him rather than the repo.

His closing line is the whole argument against the install lists, from inside the genre: "Skills are free to install. What costs you is burying the good ones under 40 that you never use."

Scan before you install — NVIDIA shipped the tool this month

There is now a scanner, it is free, and almost nobody in the install genre mentions it.

NVIDIA SkillSpector reads a skill and answers one question: is this safe to install? It carries 71 vulnerability patterns across 17 categories: prompt injection, data exfiltration, privilege escalation, supply chain, excessive agency, output handling, system prompt leakage, memory poisoning, tool misuse, rogue agent, anti-refusal, trigger abuse, dangerous code via AST, taint tracking, YARA signatures, MCP least privilege, and MCP tool poisoning.

Two things make it worth the two minutes. It runs across runtimes (Claude Code, Codex CLI, Gemini CLI and MCP skills), which matters because the SKILL.md format is shared and so are the risks. And it does two-stage analysis: fast static checks, then an optional LLM pass that compares what a skill claims to do against what its code appears to do. You can point it at a directory, a zip, a single SKILL.md or a git URL, and it emits term, JSON, markdown or SARIF, with a stable exit code you can wire into CI.

The numbers behind it, from NVIDIA's own README: 26.1% of skills contain vulnerabilities and 5.2% show likely malicious intent.

Snyk's ToxicSkills study is the other data point, and it is worth stating carefully because the headline gets miscited constantly. They scanned 3,984 skills. 1,467 had some security issue. That is where the widely-quoted 36% comes from. 534, or 13.4%, were critical: malware distribution and credential theft. And 76 were live malicious payloads aimed at OpenClaw, Claude Code and Cursor users. The 1,467 is not a payload count, however often it is repeated as one.

If you want one skill before all the others, it is Vercel's find-skills, a single SKILL.md with no bundled scripts, at 3,028,137 installs the most-installed agent skill in existence, which searches the ecosystem and ranks results by install count and source reputation rather than by whoever posted a video.

What we actually run

Briefly, because our setup is not the argument.

Our Hermes gateway has 51 skills. I wrote two of them. The rest arrived: installed once, or authored by the agent itself after it solved something twice. That ratio is the delete test failing quietly in the background, and it is why Hermes ships a Curator that archives what you stop using and a pin command for what you keep.

On the Claude Code side the number is much smaller and every one is encoded preference: the standard this article was written to, the checklist it was measured against. Those survive the delete test for the same reason Nate Herk kept his brand files: no model update will guess our house style.

If you want the security frame for the runtime rather than the packages, that is Hermes Agent Security. If you want the version of this argument about tools rather than skills, it is Codex vs Claude Code.

Where the delete advice does not apply

If you do knowledge work rather than build harnesses, do not blanket-apply it. Nate Herk makes this point against his own video's premise: the advice comes from people "designing harnesses, working on massive code bases every day, training models." If your agent formats client deliverables, the skills carrying your format are the whole value.

If a skill is a capability, deleting it just removes the capability. The six-month sweep is aimed at instruction, not at tools.

Scanning is not the same as safety. A clean SkillSpector report means 71 known patterns did not match. It does not mean the skill is trustworthy, and it does not tell you what the skill will do when a model interprets it differently next quarter.

And a caveat on the whole category. claude code skills is down about 90% over three months against explosive year-on-year growth. Some of the install-list traffic is a spike deflating, not a market forming.

The part worth keeping

The strongest thing in the interview is not the delete instruction. It is the sentence people skip past: verification is what everyone gets wrong.

Read the three sources together and they say one thing. Cherny: describe the exit criteria and let the model cook. Medin: plan the testing strategy before a line of code, so what comes back is never a first pass. Sharbel: what costs you is not the skills you install but the good ones buried under forty you never use. All three are describing a folder that gets smaller and sharper over time, not larger.

So the useful question stopped being which skills to install. It is which ones you would notice missing tomorrow. Delete the rest in six months and find out, after you have scanned whatever you are about to add.

FAQ

Should I really delete all my agent skills? Delete the ones that instruct: the ones that tell the model how to do something it already does well. Keep capability skills and keep the ones carrying context only you have. Cherny's advice is a six-month audit, not a purge.

How do I check whether a skill is safe before installing it? Run NVIDIA SkillSpector on it: point it at the directory, zip, SKILL.md or git URL. It covers 71 patterns across 17 categories and works for Claude Code, Codex CLI, Gemini CLI and MCP skills. A clean report is a floor, not a guarantee.

How many skills should I have? There is no published number, and anyone giving you one is guessing. The working test is whether you would notice a skill missing tomorrow. Both Hermes and Claude Code ship archive or curator mechanisms precisely because unused skills accumulate on their own.

Do agent skills work across different tools? Yes. SKILL.md is a shared format, adopted well beyond Anthropic. In practice the fastest path is to hand your coding agent the repository URL and ask it to install the collection for your runtime.

What is the single highest-value skill to install first? Probably find-skills from Vercel Labs, because it changes how you choose all the others: it ranks candidates by install count and source reputation instead of by who made a video about them.

---

We build AI agents and the infrastructure around them as an agency. See Next Estate, the property SaaS we run with 900+ residential complexes and 110,000+ apartments, and CashYou, a Telegram-based exchange platform we built end to end. More on what we do and who we are.