Builders mapping agent 'skills' for code
Recent creator videos show builders designing small, reusable 'skills'—constrained agent capabilities—for coding tasks like test generation, refactoring and doc extraction instead of relying on a single general assistant. The guidance emphasises explicit inputs, outputs and guardrails for each skill to improve reliability in real workflows. (youtube.com)
Coding teams are breaking one big coding agent into smaller “skills” that handle one job at a time, from refactors to test writing. (developers.openai.com) A skill is a folder with a required `SKILL.md` file and optional scripts or references, and Codex loads the full instructions only when a task matches the skill’s description. OpenAI says that “progressive disclosure” keeps context smaller by starting with just metadata such as the skill’s name and description. (developers.openai.com, agentskills.io) That pattern is now showing up in creator tutorials across tools. A January 8, 2026 Cursor video by Antoine van der Lee showed a Swift Concurrency skill used to plan and refactor a real iPhone app, and a March 2026 JetBrains video walked through creating and installing skills in Codex with modular `SKILL.md` files. (youtube.com, youtube.com) The basic idea is narrow scope. Instead of asking one assistant to “fix the codebase,” builders are writing skills for bounded tasks with clear trigger conditions, steps, and examples of what goes in and what should come out. (agentskills.io, agentskills.io) OpenAI’s Codex docs tell developers to use skills for reusable workflows and domain expertise, while keeping durable project rules in `AGENTS.md`, external system access in Model Context Protocol servers, and delegation in subagents. Anthropic’s Claude Code docs make a similar split: repeated specialist work goes to custom subagents with their own prompts, tool access, and permissions. (developers.openai.com, code.claude.com) The push toward smaller units also comes with more explicit controls. OpenAI’s Agents Software Development Kit documents separate input guardrails, output guardrails, and tool guardrails, including a blocking mode that can stop an agent before it spends tokens or runs tools. (openai.github.io) Anthropic is teaching the same operational pattern in its Claude Code training. Its “Claude Code 101” course says users can build custom subagents and skills for repeated tasks and add hooks that place deterministic guardrails around what Claude is allowed to do. (anthropic.skilljar.com) The open Agent Skills specification reflects that workflow-first design. It requires a name and description in YAML frontmatter, recommends step-by-step instructions plus input and output examples, and allows optional fields such as compatibility notes and pre-approved tools. (agentskills.io) OpenAI has also started tying skills to evaluation. A February 2026 blog post on testing Codex skills said the name and description “matter more than they might seem,” because those fields determine whether the agent loads the skill at all. (developers.openai.com) The result is less like hiring one generalist and more like keeping a shelf of checklists and scripts for recurring engineering work. The builders publishing these workflows are treating reliability as a packaging problem: define the job, limit the tools, and make the output easy to check. (developers.openai.com, agentskills.io)
