Am I using it wrong?

Detailed Analysis

A Reddit post titled "Am I using it wrong?" captures a recurring tension in the Claude Code user community between marketed capability and lived experience. The original poster describes weeks of failed attempts to build a manhwa panel-cropping tool—a computer vision task involving detecting long-stitch webtoon images, isolating individual panels, and removing speech bubbles—while comparing their plain, verbose text output to social media demos showing Opus 4.8 producing elaborate displays with visible agent orchestration and formatted tables. The poster's frustration centers on two distinct issues that often get conflated: perceived model intelligence ("Claude is dumb") and interface/output presentation (the absence of rich UI elements like tables and agent visualizations).

This complaint is significant because it highlights a gap between raw model capability and effective tool use in agentic coding workflows. Claude Code's power comes not just from the underlying LLM but from how it's prompted, scaffolded, and extended—through custom instructions (CLAUDE.md files), subagent configurations, MCP servers, and iterative task decomposition. Users who see polished demos online are often looking at curated workflows built on top of the base tool: custom slash commands, specialized subagents for planning versus execution, and carefully engineered system prompts that produce structured, formatted output. Without this scaffolding, Claude Code defaults to conversational, narrative explanations of its actions, which can feel unsophisticated even when the underlying reasoning is sound. The manhwa-cropping task itself is also nontrivial—it requires genuine computer vision engineering (edge detection, OCR or bubble-segmentation models, panel boundary heuristics) rather than pure language-model reasoning, meaning failures may reflect the inherent difficulty of the task and the need for iterative, testable subcomponents rather than a deficiency in the model.

Broader context matters here: as Anthropic and competitors push agentic coding tools toward greater autonomy, a widening skill gap has emerged between "power users" who treat these tools as configurable development environments and casual users who treat them as chatbots. The former group invests in prompt engineering, project-specific context files, breaking tasks into verifiable increments, and using planning-mode or extended-thinking features before execution—practices that dramatically change output quality and format. This mirrors a pattern seen throughout the LLM era, from prompt-engineering communities around GPT models to today's proliferation of Claude Code "cookbooks," custom hooks, and agent frameworks. The perception that a top-tier model is "dumb" often traces back not to model regression but to insufficient context, unclear task decomposition, or unrealistic expectations set by highlight-reel demonstrations that omit the extensive setup behind them.

Ultimately, this thread reflects a growing need for better onboarding and documentation around agentic coding tools as they become more powerful and more widely adopted. Anthropic has increasingly published guidance on effective Claude Code usage—covering CLAUDE.md conventions, subagent design, and iterative workflows—precisely because the gap between default behavior and expert-configured behavior is substantial. As AI coding assistants continue to advance in raw capability, the differentiator for real-world outcomes is shifting toward user skill in context engineering, task scoping, and tool configuration, suggesting that "prompting literacy" is becoming as important to developer productivity as the models themselves.