LLMs struggle to replicate the flow state of handwritten coding, frustrating developers who see AI as a productivity tool. A Hacker News discussion highlights how current tools like Claude Code and Codex disrupt workflows with constant prompt-response cycles. Users report stuttering interactions where models require frequent re-prompting, creating a "bicycle that brakes abruptly" effect. The core issue? Over-reliance on external orchestration instead of letting models handle complexity.

The thread proposes a tab model as a potential alternative, where context is loaded lazily and skills—rather than agents—drive capabilities. One contributor argues skills should act as custom instructions for tasks outside a model's training data, like interfacing with proprietary systems. They emphasize clean contexts and precise specifications, letting models self-verify goals without verbose scaffolding. This approach contrasts with conventional advice, which piles on memory and hooks, risking "context rot" and diluted intent. The debate centers on whether skills expand what LLMs can do or if agents merely manage context.

A key takeaway is the undervalued power of skills. Unlike agents, which focus on maintaining context, skills enable models to execute specific, non-standard tasks. For example, embedding code in skills allows targeted actions without bloating context. The author prefers a "clean slate" approach, where models only access necessary information when needed—akin to progressive disclosure in engineering. This philosophy challenges developers to prioritize clarity in specs and exit criteria, reducing technical debt from over-engineered orchestration layers. The discussion reflects a broader shift in how programmers interact with AI, moving from trial-and-error prompting to structured, skill-driven workflows.

While the thread lacks concrete implementations, it underscores a critical question: Can LLMs evolve beyond the prompt-response loop to enable seamless coding? The emphasis on skills suggests yes, but practical adoption remains unproven. Startups or researchers experimenting with tab models or skill-based architectures could redefine AI-assisted development. Until then, developers like the original poster remain stuck in a cycle of fragmented interactions. The debate highlights a tension between current tooling limitations and the theoretical potential of LLMs as true cognitive amplifiers.