Coding agents need 'skills' not prompts
Recent creator guides argue that AI coding agents are becoming more effective when built as composable ‘skills’—read, change, verify, explain and escalate—rather than as ad‑hoc LLM prompts. (youtube.com)
Coding agents work better when teams package repeatable jobs as reusable “skills” instead of rewriting long prompts for every task. (developers.openai.com) A skill is a folder with a `SKILL.md` file and optional scripts or reference files, and Codex can load it in the command line tool, integrated development environment extension, and Codex app. Codex starts with a skill’s metadata and loads the full instructions only when the task matches, which keeps context smaller than pasting the whole workflow into every prompt. (developers.openai.com) Microsoft describes Agent Skills the same way in Visual Studio Code: reusable folders of instructions, scripts, and resources that GitHub Copilot can discover automatically and load only when relevant. Its documentation says skills are for specialized workflows such as testing, debugging, and deployment, while custom instructions are for standing coding rules. (code.visualstudio.com) That is a shift from the earlier habit of “prompt engineering,” where developers tried to cram every rule into one message and hoped the model would remember it. GitHub’s Copilot documentation says skills replace older `*.prompt.md` files because skills can be discovered by agents automatically and can carry scripts, templates, and other assets alongside instructions. (awesome-copilot.github.com) OpenAI’s Codex guidance now tells users to treat a coding agent less like a one-off assistant and more like “a teammate you configure and improve over time.” The same guide says repeated work should be turned into skills, while durable repository-wide guidance belongs in files such as `AGENTS.md`. (developers.openai.com) GitHub added Agent Skills support to Copilot on December 18, 2025, and said skills work across Copilot coding agent, Copilot command line interface, and agent mode in Visual Studio Code Insiders. GitHub also said repositories that already used Claude Code’s `.claude/skills` directory could be picked up automatically by Copilot. (github.blog) Anthropic has pushed the same pattern in Claude. Its public `anthropics/skills` repository says skills are folders of instructions, scripts, and resources that Claude loads dynamically for specialized tasks, and Anthropic’s product blog says skills are composable and can be used across Claude apps, Claude Code, and the application programming interface. (github.com, claude.com) OpenAI, Microsoft, and Anthropic are all converging on the same basic architecture: keep general reasoning in the model, but store repeatable engineering habits outside the prompt as named modules. In practice, those modules can cover steps like planning, making code changes, running tests, reviewing output, and deciding when to ask a human for help. (developers.openai.com, code.visualstudio.com, github.com) The open-source ecosystem has moved fast around that idea. Addy Osmani’s `agent-skills` repository, which packages workflows such as `/spec`, `/plan`, `/build`, `/test`, `/review`, and `/ship`, had about 11,700 GitHub stars and 1,300 forks as of April 16, 2026. (github.com) The appeal is less about giving a model clever words and more about giving it a checklist with tools attached. As coding agents spread from chat boxes into editors and command lines, the winning pattern looks closer to software packaging than prompt writing. (developers.openai.com, developers.openai.com)
