From messy ideas to working agents in a single session
Most developers think of Claude Code as a coding assistant. You point it at a codebase, describe what you want, and it writes code. I've been using it for something different: rapidly prototyping agentic workflows that have nothing to do with software development.
This started naturally. I had a repetitive research task and figured I'd try describing it to Claude Code instead of writing a script. It worked surprisingly well. Turns out I'd stumbled onto something the Claude Code team already knew. As Cat Wu, Claude Code's Product Manager at Anthropic, put it in a recent conversation: "the underlying primitives were really general" and "if you squint, you can kind of build anything on it." At its core, Claude Code is an agent that can run (quasi) indefinitely on a concrete task, access the internet, write and execute code, and manage files. That's not a coding tool. That's a general-purpose agent harness.
Zero to Prototype
The friction of building agents often kills good ideas before they get tested. You have a workflow in mind—something too messy for traditional automation but too tedious to do manually every time—and by the time you've configured an agent framework, written the glue code, and debugged the first run, you've lost momentum. Claude Code eliminates that friction. You can go from idea to working prototype in a single session, iterating through conversation until the workflow does what you need.
This is what makes it valuable for prototyping agentic workflows: research tasks, document processing, data cleaning, content generation pipelines. The kind of work where you know roughly what you want but can't specify it precisely upfront. Claude Code lets you explore the problem space interactively, refining your approach until it works, before committing to any framework.
Agent Primitives, No Assembly Required
Claude Code gives you access to several primitives out of the box: file operations, shell commands, web search, code execution, and (via MCP) extensibility to external services. These are the same building blocks you'd wire together in a formal agent framework. The difference is you can interact with them conversationally, iterate in real time, and test ideas in minutes rather than days.
Consider a task like processing expense reports. With Claude Code, I can describe the workflow: "Pull transactions from this CSV, categorize them based on merchant names, check against our expense policy, and flag anything that needs review." Claude Code can read the file, write Python to parse and categorize, execute it, and show me results. I can refine the categorization logic through conversation until it works. No agent framework setup, no LLM API token, no boilerplate.
From Conversation to Slash Command
The pattern I've landed on follows a few steps. First, gather your source materials in local folders. Don't overthink the structure. Then start an interactive session and describe your goal. Let Claude Code explore different approaches. When something works well, use /rewind to clean up false starts and convert the conversation into a slash command, a reusable markdown file capturing the logic and guidance.
The mapping to agentic workflows is surprisingly direct. Slash commands become your workflow entry points—reusable prompts that kick off a defined process. Sub-agents handle specialized tasks in parallel, just like agents in a multi-agent system. The orchestration happens through conversation rather than code, but the structure is the same.
This is the key insight from Matt Stockton's write-up on the approach: the work isn't throwaway. The interactive prototyping becomes the foundation for automation later. You can port a working slash command into the Claude Agent SDK or adapt it to other frameworks with the hard problems already solved.
When to Graduate
Once a workflow is validated through interactive prototyping, you have options. The Claude Agent SDK, which powers Claude Code itself, provides the infrastructure for production agents. Or you can take what you've learned to frameworks like CrewAI or LangGraph. You might also swap Claude Code's agentic search for a specialized RAG system like Contextual AI when retrieval quality matters. The point is you're not starting from scratch—you've already solved the hard problems.
The value of starting with Claude Code is you've already answered the hard questions. Does the approach work at all? What edge cases matter? What does good output look like? You bring those answers, and often the actual prompts and logic, into whatever production framework you choose.
For anyone building agents, I'd recommend trying this before reaching for a framework. Spin up Claude Code, describe a workflow you've been meaning to automate, and see how far you get in an hour. The primitives are more general than the name suggests.
Joe Marlo
Director of Data Science
Lander Analytics
Subscribe to our Substack and below to our monthly emails for practical AI strategies for your organization: what to build, what to avoid, and how to make systems reliable in the real world.
Work with us: If you want help identifying the right first workflow, building a permissioned knowledge base, or training your team to ship responsibly, reach out at info@landeranalytics.com.
About the author: Joe Marlo is Director of Data Science at Lander Analytics, where he designs agentic workflows, statistical models, and interactive frontends that put rigorous analysis into production.
