For over twenty years, Amazon’s Working Backwards process has helped drive customer focus and clarity of thought. We start by writing a press release and FAQ before we build, ensuring we think deeply about the customer problem and solution. But sometimes we know the problem without yet knowing the right solution or whether our ideas are technically feasible. In these ambiguous spaces, AI enables a new approach: prototyping before writing the PRFAQ.
In the past, building even a rough prototype could take months of coding. The high up-front cost led us to skip prototyping and jump straight into writing a PRFAQ, trying to describe something we’d never seen or used. Large language models and agentic coding have changed this. We can now build working prototypes in days instead of months, helping us refine ideas and vet feasibility before starting a PRFAQ.
Early in 2023, a small group and I started brainstorming ideas for how Generative AI could help us make AWS easier to use. We didn’t know what the models were capable of, and we didn’t know if our ideas were feasible. So we just tried it out. In a few days, we had a working prototype that could analyze error messages from different AWS services and identify their root causes. This eventually became “Diagnose with Amazon Q.”
Another developer built an internal LLM playground as a side project. It went viral with Amazonians, revealing a problem we weren’t aware of—that our builders wanted an easy way to do hands-on experimentation with Generative AI. That prototype inspired PartyRock.aws, and since its public launch, over half a million apps have been built by users worldwide.
The genie was out of the bottle. Our imaginations had been sparked, and we had an idea of what we could actually do with this technology.
Last year, when we set out to build an agentic IDE focused on spec-driven development (which would become Kiro), we couldn’t have written a PRFAQ at the outset. We had no idea what it would look like, but we wanted to help developers be more effective at agentic coding. Prototyping let us try out a bunch of ideas and actually use them in our daily work. The Kiro team started using the IDE full time as soon as we built the first working prototype. They could then evolve it from that first prototype into an agentic IDE they loved using.
Our Working Backwards process remains as relevant as ever. But in the age of AI, sometimes the fastest path to clarity is a quick prototype. Prototypes built with AI give us a powerful tool for exploring ambiguous spaces and validating technical feasibility before we commit to the full rigor of a PRFAQ.