“Skills” Are Not Software Engineering
Anthropic’s “skills” framework [1] is a useful step toward more capable AI agents. Skills package instructions, resources, and executable scripts into reusable tools that Claude can call to perform actions across contexts. This makes advanced workflows accessible to users who are not software engineers, and it lets LLMs offload deterministic execution to real code. Skills clearly improve reliability compared to pure prompting.
At the same time, skills cannot replace software engineering. They help with local execution, but they do not solve the broader orchestration and correctness problems that real software systems must address. Skills still rely on natural language as the primary interface for interpreting intent, selecting tools, filling parameters, and deciding sequences of actions. Natural language is valuable because it allows the model to apply commonsense reasoning and to bridge ambiguity, but it also introduces ambiguity of its own, since there are no static contracts, type checks, or guarantees that the same request will lead to the same plan over time.
Scripted skill bodies execute deterministically, but everything around them remains probabilistic. The model decides when to invoke a skill, how to connect skills together, how to interpret return values, and how to recover from errors. These decisions rely on model reasoning rather than explicit control flow, which means silent misrouting, partial failures, and unpredictable drift can still occur. Software engineering requires predictable interfaces, state management, transactions, testing, and clear error semantics. Skills do not yet provide these foundations.
A more reliable path forward combines the strengths of LLMs with the rigor of traditional engineering. Natural language should remain the medium for planning, intent understanding, disambiguation, and high-level reasoning. A deterministic runtime should then validate and compile that plan into a typed, structured workflow that enforces schemas, invariants, contracts, and version guarantees. In this hybrid model, the LLM proposes, while the runtime ensures correctness.
Anthropic’s skills are a meaningful contribution to agent tooling. They expand what non-engineers can do and make LLM-driven workflows more powerful. But until skills are paired with a structured execution layer that enforces engineering discipline, they remain an augmentation of prompting rather than a foundation for building reliable systems. What we need next is a new software engineering paradigm for the LLM era (e.g., [2][3][4][5]), one that preserves the flexibility and reasoning strengths of natural language while grounding execution in typed interfaces, verifiable workflows, and deterministic runtimes. This hybrid model will allow us to embrace LLM agency without sacrificing the predictability, safety, and maintainability that real software demands.
References
[1] Anthropic. October 16, 2025. “Introducing Agent Skills.” https://www.claude.com/blog/skills
[2] Honghua Dong, Qidong Su, Yubo Gao, Zhaoyu Li, Yangjun Ruan, Gennady Pekhimenko, Chris J. Maddison, and Xujie Si. 2024. “APPL: A Prompt Programming Language for Harmonious Integration of Programs and Large Language Model Prompts.” https://arxiv.org/abs/2406.13161
[3] Jayanaka L. Dantanarayana, Yiping Kang, Kugesan Sivasothynathan, Christopher Clarke, Baichuan Li, Savini Kashmira, Krisztian Flautner, Lingjia Tang, and Jason Mars. 2024. “MTP: A Meaning-Typed Language Abstraction for AI-Integrated Programming.” https://arxiv.org/abs/2405.08965
[4] Stephen Mell, Konstantinos Kallas, Steve Zdancewic, and Osbert Bastani. 2024. “Opportunistically Parallel Lambda Calculus.” https://arxiv.org/abs/2405.11361
[5] Eagon Meng and Daniel Jackson. 2025. “What you see is what it does: A structural pattern for legible software.” https://arxiv.org/abs/2508.14511
Originally published on LinkedIn.