Derive is building an open repository for AI workflows and
If you are currently managing your prompt engineering through messy Notion pages or random text files, you know how quickly things fall apart when you try to replicate a specific result across different models. Derive aims to solve this by providing a structured way to host workflows. Instead of just sharing a single prompt, you can share the entire context: the system instructions, the temperature settings, the few-shot examples, and the specific model parameters that made the result work.
How the workflow management works
The platform is designed to treat an AI interaction not as a one-off event, but as a repeatable asset. When you build a workflow on Derive, you aren't just saving text; you are documenting an entire execution path.
1. Artifact Creation: You define the core components of your task, ranging from simple text transformations to complex multi-step reasoning chains.
2. Version Control: As you tweak your prompt engineering to handle edge cases, you can iterate on the version without losing the original baseline that worked.
3. Public/Private Sharing: You can host your workflows openly for the community to fork or keep them private for internal company use.
The real value here is the ability to create a "hands-on guide" for others to follow. If I develop a particularly effective way to extract structured JSON from messy legal documents using a specific Claude 3.5 Sonnet configuration, I can upload that entire workflow to Derive. Someone else can then pull that workflow, see exactly how the prompt was structured, and deploy it into their own AI workflow immediately.
Why this matters for LLM agents
As we move away from simple chatbots and toward autonomous LLM agents, the complexity of what we are "programming" increases exponentially. We aren't just writing code anymore; we are orchestrating chains of thought and tool-calling sequences.
Managing these sequences requires more than just a copy-paste buffer. We need a way to inspect the intermediate steps of an agent's logic. Derive provides that infrastructure. By treating these sequences as first-class citizens—artifacts that can be inspected, tested, and versioned—it bridges the gap between "playing with prompts" and actual production-grade AI deployment.
It’s essentially a move toward making prompt engineering more scientific and less experimental. If you are building complex agentic systems, having a dedicated home for your logic is going to be a requirement, not a luxury.