Detailed Analysis
A Reddit post from a self-described "staff level engineer" offers a granular, practitioner-level account of building a health-focused IP product using a multi-tool AI workflow that leans heavily on Claude's Opus model alongside a tool called Fable. The author's central claim is counterintuitive: Fable, often evaluated purely as a coding assistant, is "just ok" at writing code but excels as a planning and architecture partner—functioning less like an autocomplete engine and more like an "assistant engineering manager." Meanwhile, Claude Opus is credited with handling more direct coding tasks, such as building out a design system and largely autonomously implementing a marketing site with minimal oversight. This division of labor—Claude for execution, Fable for planning and rigor—reflects a maturing pattern among technically sophisticated users who are learning to route different parts of the software lifecycle to differently-strengthed AI tools rather than treating any single model as a one-size-fits-all solution.
The specifics of the workflow are notable for how they illustrate real engineering discipline being layered onto AI assistance rather than replaced by it. The author describes using Fable to rigorously plan API and middleware architecture for a non-monolithic system, resulting in concrete technical decisions: drift checks to catch sync failures between middleware and consumer layers, and a "moat" architecture that siloes IP-sensitive backend logic behind public IDs safe for browser exposure. These are not trivial scaffolding choices—they reflect genuine software architecture concerns around data protection and system integrity, suggesting that the AI tooling was pushed to reason about second-order engineering risks (security, consistency, technical debt) rather than just producing working code. The final production pattern—Fable for planning and test-swarming, Opus for code generation, and GitHub Copilot for human-supervised review—represents a triangulated AI pipeline where each tool's output is checked by another, effectively creating a lightweight multi-agent code review process without formal orchestration infrastructure.
This account matters within the broader Claude ecosystem because it exemplifies a shift in how professional engineers talk about AI coding assistants: less in terms of raw code-generation benchmarks and more in terms of workflow orchestration, architectural reasoning, and risk mitigation. Claude's Opus model has increasingly been positioned by users and Anthropic alike as strong not just at writing syntactically correct code but at handling large-context reasoning tasks—here, autonomously building a marketing site "with little massaging," which speaks to growing trust in longer unsupervised runs for well-scoped tasks. The author's emphasis on protecting proprietary IP while using AI tools also touches a live concern in the developer community: how to leverage powerful, cloud-based coding assistants without leaking sensitive business logic or data, a tension addressed here through architectural siloing rather than avoidance of AI tools altogether.
More broadly, this post reflects a trend of solo or small-team builders using layered AI assistance to compress timelines that would traditionally require a full engineering team—describing an API layer built in days and a full application in weeks. It also signals the emergence of a "planning-first" AI usage pattern, where tools are evaluated not just on code quality but on their capacity to surface architectural decisions a human might have made anyway, only faster and with fewer blind spots. As AI coding tools proliferate and differentiate—Claude, Fable, Copilot, and others—users are increasingly building bespoke pipelines that treat each tool as a specialized collaborator with distinct strengths, foreshadowing a future where "which AI is best at coding" becomes a less relevant question than "which AI is best at which part of the engineering process."
Read original article →