---
title: "I Just Deleted All My Agent Skills"
description: "Do agent skills actually work? Paired benchmarks of public skills show little gain and higher token costs, so I deleted mine. See where the data disagrees."
date: 2026-11-07
author: "Philipp D. Dubach"
categories:
  - "AI"
keywords:
  - "do agent skills work"
  - "agent skills benchmark"
  - "are Claude Code skills worth it"
  - "delete Claude Code skills"
  - "agent skills security"
  - "agent skills"
  - "Claude Code skills"
  - "SKILL.md"
  - "SWE-Skills-Bench"
  - "SkillsBench"
  - "curated agent skills"
  - "agent skills token cost"
  - "AGENTS.md study"
  - "CLAUDE.md"
  - "coding agent context files"
  - "malicious agent skills"
  - "context engineering"
  - "coding agents"
type: "Commentary"
canonical_url: "https://philippdubach.com/posts/i-just-deleted-all-my-agent-skills/"
source_url: "https://philippdubach.com/posts/i-just-deleted-all-my-agent-skills/index.md"
---

# I Just Deleted All My Agent Skills

*Philipp D. Dubach · Published November 7, 2026*


## Key Takeaways

- In one paired benchmark, 92 of 98 public software engineering skills had no significant effect on coding task results.
- A web development benchmark found that public skills lowered mean pass rates and raised token cost by 72% to 394%.
- Curated skills raised average pass rates from 33.9% to 50.5% in SkillsBench, so the evidence does not condemn every skill.


---


I just deleted all my agent skills after reading [Jakob Breu's](https://jakobbr.eu/2026/10/14/hot-take-ai-skills-already-getting-obsolete-after-months/) (hot) take. I kind of had a gut feeling myself for a long time as models evolve but I didn't really have the evidence to back it up, so I went looking for it. Some of what I found:

- [SWE-Skills-Bench](https://arxiv.org/abs/2603.15401) tested 98 public skills. 92 had no significant effect, and the average pass ratio went from 76.6% to 76.9%.
- [Web development skills](https://arxiv.org/abs/2608.23067) made results slightly worse and cost 72% to 394% more tokens.
- [AGENTS.md files](https://arxiv.org/abs/2602.11988) fare little better: no general gain in task success, over 20% more cost.

Also, Anthropic [cut over 80% of Claude Code's system prompt](https://claude.dev/blog/the-new-rules-of-context-engineering-for-claude-5-generation-models/) with "no measurable loss" on its own evaluations, and says it was overconstraining the agent, skills included. [Karpathy's prompts keep getting shorter](https://news.ycombinator.com/item?id=50323111). Boris Cherny [reportedly says](https://charliehills.substack.com/p/delete-your-claudemd) to delete the lot every six months.

Then there's security. There seems to be at least one [vulnerability in about a quarter](https://arxiv.org/abs/2601.10338) of 31,132 skills. [A second study](https://arxiv.org/abs/2602.06547) found similar problems.

*Related: [Karpathy's Software 3.0 Playbook](https://philippdubach.com/posts/karpathys-software-3.0-playbook/)*

I might regret some of this. [SkillsBench](https://arxiv.org/abs/2602.12670) found curated skills lifted pass rates from 33.9% to 50.5%. [Jim Bennett's line](https://jimbobbennett.dev/blogs/skills-are-pipelines/) seems right to me: skills that repeat the docs can go, but no model can infer your own process and voice.

Anthropic says much the same, as I understand it.

---

## Frequently Asked Questions


### Do agent skills actually work?

Public skills mostly do not. SWE-Skills-Bench tested 98 public skills and found no significant effect for 92 of them, and the average pass ratio rose from 76.6% to 76.9%. Curated skills did better in SkillsBench, where the average pass rate rose from 33.9% to 50.5%.


### Do agent skills increase token costs?

They can. In one web development benchmark, public skills cost 72% to 394% more tokens and made results slightly worse. AGENTS.md files added over 20% to cost with no general gain in task success.


### Which Claude Code skills are worth keeping?

Keep skills that hold knowledge a model cannot infer, such as your own processes, rules or voice. Anthropic says skills work best when they encode knowledge particular to you, your team or your product. Skills that explain a documented tool are easier to remove and test.


### Are agent skills safe to install?

Public skills carry security risk. One scan found at least one vulnerability in about a quarter of 31,132 skills.



---

Canonical: https://philippdubach.com/posts/i-just-deleted-all-my-agent-skills/
This file is the canonical machine-readable variant of https://philippdubach.com/posts/i-just-deleted-all-my-agent-skills/. Author: Philipp D. Dubach (https://philippdubach.com/).