Claude skills: start lean with agents
Greg Isenberg highlighted a lean approach to building Claude skills and AI agents—favour tested, small skills over bloated setups, and build recursively with strong context windows. The recommendation is to keep agent stacks simple and reliable rather than packing in unproven features. (x.com)
Most people building an artificial intelligence agent start by bolting on memory, tools, prompts, and five different automations at once. Anthropic’s own setup for Claude points the other way: a “skill” is just a small folder of instructions, metadata, and optional scripts that Claude loads when the task matches. (platform.claude.com) That sounds simple because it is simple. Anthropic says skills are for repeatable jobs like analyzing data with one company workflow or creating documents with one brand style, not for turning every workflow into a giant operating system. (platform.claude.com, resources.anthropic.com) Greg Isenberg has been pushing that same pattern in recent Claude tutorials and podcast episodes. His examples center on narrow jobs like copy review, tweet-to-newsletter conversion, and other single-purpose helpers that can be tested quickly instead of “agent armies” built all at once. (podcasts.apple.com, podwise.ai, pod.wave.co) The reason small skills beat bloated agents is context. Every extra instruction, tool description, and memory file competes for space inside the model’s working memory, which is the text window Claude has to read before it answers. (platform.claude.com, code.claude.com) Anthropic’s answer to that problem is separation. In Claude Code, a subagent runs in its own context window with its own system prompt, tool access, and permissions, so one specialist can handle one job without dragging the whole conversation history into every step. (code.claude.com) That is the “build recursively” idea in plain English. Instead of one giant assistant that tries to remember your marketing playbook, coding rules, research method, and design style at the same time, you make a small specialist for one task and let another specialist call it when needed. (code.claude.com, platform.claude.com) Anthropic’s public skills repository shows what that looks like in practice. The company describes skills as folders of instructions, scripts, and resources that Claude loads dynamically for specialized tasks, which is much closer to a toolbox of tested jigs than a single all-knowing robot employee. (github.com) The same pattern shows up in Anthropic’s newer managed-agents work. The company says the stable pieces are the session, the harness that routes tool calls, and the sandbox where code runs, while the implementation underneath can change without breaking the whole system. (anthropic.com) That is why the lean advice keeps coming up: keep the abstractions steady and the parts swappable. A small skill that reliably writes one kind of report is easier to debug, measure, and reuse than a sprawling agent stack that sometimes researches, sometimes edits, and sometimes fails because one tool changed. (anthropic.com, resources.anthropic.com) So the practical playbook is not “add more agent features.” It is “pick one repeated task, make one skill, test it on real work, and only then let other skills or subagents call it,” which is exactly how Claude’s own docs frame modular capability instead of one giant prompt. (platform.claude.com, code.claude.com)
