A developer just Deleted All My Agent Skills

Software developer Philipp Dubach purged his entire repository of agent skills after reviewing empirical research and an industry critique from Jakob Breu. As detailed in philippdubach.com's report, Dubach verified his intuition by examining evaluation benchmarks, finding that most public assistant extensions fail to enhance performance while creating latency, token consumption, and security risks.

This report examines benchmark findings, prompt reduction strategies from leading artificial intelligence labs, and pass rates across public and curated extensions. The analysis provides crucial context for software engineers and architects managing autonomous coding agents, detailing token overhead, context bloat, vulnerability distributions, and the narrow boundaries where specialized instructions succeed.

A developer just Deleted All My Agent Skills: what actually changed

The decision to remove all agent skills followed a rigorous assessment of empirical data rather than speculative opinion. Looking into benchmark evidence after reading an industry critique from Jakob Breu, the investigation uncovered that public coding extensions provide virtually no measurable benefit to underlying models. In the paired benchmark SWE-Skills-Bench, researchers evaluated ninety-eight public skills across standardized coding tasks. Ninety-two of those ninety-eight skills showed no statistically significant effect on task completion, shifting the average pass ratio merely from seventy-six point six percent to seventy-six point nine percent.

Beyond general coding tasks, web development extensions demonstrated an actively detrimental impact on system performance. Across a dedicated web development benchmark, public skills lowered mean pass rates while drastically inflating operational expense, driving token usage up between seventy-two percent and three hundred ninety-four percent. Repository configuration files fared no better under scrutiny. Testing revealed that AGENTS.md files produced no general gain in overall task success while increasing context token costs by over twenty percent, exposing widespread operational inefficiency across common agent tooling setups.

A developer just Deleted All My Agent Skills: how it works

The operational degradation caused by public agent skills stems from overconstraining the foundational reasoning mechanisms of large language models. Leading frontier model developers have observed that excessive instructional scaffolding frequently degrades autonomous behavior. Anthropic removed more than eighty percent of the Claude Code system prompt and recorded no measurable loss across its internal evaluation suites. Anthropic stated that the previous verbose instructions, including bundled skills, actively overconstrained the agent during execution, demonstrating that large context prompts restrict flexible problem-solving routines on practical tasks.

Industry practitioners have embraced a similar discipline of aggressive prompt simplification. Computer scientist Andrej Karpathy has steadily shortened his system prompts as underlying foundation models advanced in reasoning capability. Boris Cherny reportedly advocates deleting an entire agent configuration setup every six months to eliminate obsolete instructions. Public agent skills frequently attempt to explain basic library behaviors that modern models already understand, burdening the context window with redundant documentation rather than allowing the model to apply its native technical knowledge directly to programming tasks.

A developer just Deleted All My Agent Skills: why it matters

The widespread deployment of unverified public skills introduces severe security risks alongside performance penalties. An audit of thirty-one thousand one hundred thirty-two public agent skills revealed that approximately one-quarter of the examined catalog contained at least one security vulnerability. A second independent study confirmed similar vulnerability rates across public skill repositories. Integrating external third-party instructions directly into agent runtime environments exposes developer machines, internal networks, and software repositories to potential compromise, transforming unvetted instructional packages into significant security liabilities for development organizations.

The operational cost of maintaining redundant skills is equally unsustainable. When web development skills elevate token consumption by seventy-two percent to three hundred ninety-four percent without improving output quality, engineering teams incur substantial financial costs for degraded results. Similarly, maintaining AGENTS.md configurations that inflate execution costs by more than twenty percent without boosting task completion drains computing resources. Eliminating unnecessary public skills immediately reduces API billing, removes fragile context dependencies, and eliminates unvetted third-party instructional payloads from automated software development workflows.

A developer just Deleted All My Agent Skills: who is affected

This transition directly impacts software engineers and engineering teams relying on public skill registries to augment autonomous coding workflows. Developers who assumed that accumulating community-authored agent extensions would steadily improve coding performance face clear benchmark data indicating that ninety-two out of ninety-eight public skills produce no meaningful effect. Organizations maintaining large shared AGENTS.md repositories are also affected, as data proves these configuration files increase token billing by more than twenty percent while providing no overall improvement in task completion rates.

Engineering managers and security teams responsible for enterprise artificial intelligence governance must immediately review the security exposure of automated agents. With roughly twenty-five percent of more than thirty-one thousand public skills harboring security flaws, development teams that permit unvetted skills in local development environments risk enterprise safety. Technical leads must differentiate between redundant public extensions that replicate standard API documentation and carefully governed internal rules that encode necessary private operational standards across their ongoing software development lifecycles.

A developer just Deleted All My Agent Skills: what to watch

The empirical evidence does not mandate the total abandonment of all agent skills, establishing a sharp distinction between public documentation scrapers and carefully curated workflow guides. In evaluations conducted on SkillsBench, curated skills raised average pass rates from thirty-three point nine percent to fifty point five percent, proving that specialized instruction sets can deliver measurable improvements. As noted by Jim Bennett, skills that merely repeat public software documentation can be deleted safely, but foundational models cannot infer an individual team's private engineering processes, specialized domain rules, or organizational voice.

Anthropic holds a similar position regarding the boundary between generic prompt stuffing and essential context guidance. Moving forward, engineering organizations should monitor automated evaluations, establish scheduled six-month purges of legacy agent prompts, and restrict custom skills to proprietary architectures and organizational practices. Stripping away public extensions that merely restate public manuals leaves developers with lean, cost-efficient agents that leverage pure model intelligence while applying targeted, curated rules exclusively where proprietary operational logic requires explicit instruction and guidance.