[{"data":1,"prerenderedAt":19},["ShallowReactive",2],{"post-a-day-with-fable-mythos":3},{"excerpt":4,"postId":5,"status":6,"tags":7,"title":12,"publishedAt":13,"updatedAt":14,"coverImage":15,"slug":16,"content":17,"createdAt":18},"What changed my thinking was not the code it produced. It was how much of the workflow it understood on its own.","e5e9410f-9dab-458a-b752-2cc63c52108f","published",[8,9,10,11],"Fable","Mythos","Claude Code","AI Workflow","A Day With Fable / Mythos","2026-06-29T19:43:57.197Z","2026-06-29T19:54:40.832Z","https://ericnparadis-com-prod-mediabucketbucket-baexchhz.s3.amazonaws.com/media/6c35a44e-1916-4d57-82f1-9fa6c447799c.png","a-day-with-fable-mythos","I had a chance to spend a day working with Fable / Mythos, and it left me thinking less about that specific model and more about how quickly the AI development experience is changing.\n\nThe work was for OnBrandify, specifically around variable print-on-demand template creation and customer order customization. It is the kind of product work that gets complicated quickly. You are not just building a form or a checkout flow. You are dealing with brand-controlled templates, editable fields, customer-specific customization, preview accuracy, production rules, approval logic, and the handoff from a digital order to something that can actually be printed correctly.\n\nThat kind of workflow has a lot of hidden complexity. A template needs to be flexible enough for the customer to personalize, but controlled enough that the brand does not lose consistency. The order experience needs to feel simple, even though there may be a lot happening underneath it.\n\nNormally, getting AI to help with that kind of work takes a fair amount of setup. You need to explain the product, the architecture, the patterns in the codebase, and the decisions you want it to make. You often need to break the work into smaller pieces and keep pulling the model back when it starts solving the wrong part of the problem.\n\nThis felt different.\n\nWith a pretty basic prompt and not much context beyond the existing codebase, the model did not just jump into writing code. It asked good clarifying questions. It explored the problem from multiple angles. It came back with a thoughtful plan that was more strategic than tactical, and it asked for input before moving forward.\n\nThen it proposed how to actually do the work. It separated the effort into discovery, implementation, testing, and verification. It watched the CI pipeline, corrected errors, and kept looping through the work until the tests and deployment checks were passing.\n\nThat was the part that stuck with me.\n\nThe impressive thing was not that it could generate code. That is becoming less surprising every month. The impressive thing was that it seemed to understand the shape of the work. It recognized that this was not just a coding task. It was a product and workflow problem that needed to be understood before it could be built well.\n\nRight now, a lot of effective AI work still depends on knowing how to set up the model properly. You need good prompts. You need the right context. You need clear instructions. You need to know when to split work into separate tasks, when to slow the model down, and when to verify what it produced.\n\nIn other words, to get great results from AI, you often still need to be fairly good at working with AI.\n\nBut experiences like this make me wonder how long that remains true.\n\nThe better these models get, the more the structure starts to move into the model itself. It can ask for missing context. It can propose a plan. It can recognize the steps needed to get from idea to working software. It can handle more of the back-and-forth that used to require a person to carefully manage the process.\n\nThat does not mean prompts, context, or workflow design stop mattering. They still do. But it does suggest that some of the AI-specific expertise we have been building around the models may become less visible over time.\n\nThe advantage starts to shift.\n\nToday, a lot of the advantage comes from knowing how to operate the AI machinery. You get better results because you know how to structure the task, guide the model, and keep the work on track.\n\nBut as the models get better, more of the advantage moves toward knowing what good work looks like.\n\nAnd that raises the more interesting question for me: what is the human role as more of the workflow moves into the model?\n\nIs it product judgment?\n\nTechnical judgment?\n\nUser experience?\n\nCustomer experience design?\n\nTaste?\n\nMaybe it is all of those things. The ability to look at the work and know whether the plan is sound, whether the flow makes sense, whether the implementation fits the product, and whether the output is accurate enough for real customers.\n\nThat is what made the OnBrandify experience feel different. It was not just that AI made the work faster. It reduced the amount of AI-specific setup required to get to a thoughtful, sophisticated workflow.\n\nFor a product where the goal is to make complex co-branded marketing and print customization feel simple to the customer, that feels especially relevant. The work underneath may still be complicated. But the experience on top gets smoother.\n\nThat may be the bigger pattern with AI as well.\n\nThe expertise does not disappear.\n\nIt moves up the stack.","2026-06-29T19:43:57.199Z",1782763014528]