Detailed Analysis
Anthropic's Claude Code, the company's agentic coding tool designed to integrate directly with developers' editors and terminals, has drawn attention for a specific default behavior: automatically attaching the currently open file in a user's IDE to every prompt sent to the model. The article's framing—"Don't Attach Open File by Default"—reflects a piece of practical developer feedback suggesting that this convenience feature, while well-intentioned, can create friction in real-world workflows. When a tool silently injects file context that a developer didn't explicitly request, it can lead to unexpected token consumption, diluted or irrelevant context being fed to the model, and confusion when the assistant's responses seem to reference files the user wasn't actually asking about.
This detail matters because it touches on a broader design tension in agentic coding tools: the balance between automatic context-gathering and user control. Claude Code, like competitors such as GitHub Copilot Workspace, Cursor, and Cody, tries to reduce the friction of manually specifying which files are relevant to a coding task. Automatically attaching open files is a reasonable heuristic—most developers are looking at the file they care about—but heuristics like this can misfire, especially in multi-file projects, when switching between tabs quickly, or when a developer has a file open for reference rather than for editing. The complaint underscores that even subtle defaults in AI coding assistants can have outsized effects on cost (since context window usage often ties directly to API billing), latency, and the quality of generated suggestions.
The issue also reflects a maturing phase in the agentic coding tool market, where early adopters are moving past marveling at raw capability and are now scrutinizing the granular UX decisions that determine whether these tools fit into disciplined engineering workflows. As coding agents become more autonomous—executing multi-step tasks, editing multiple files, and running terminal commands—the question of what context they should have access to, and when, becomes increasingly important. Overly aggressive default context injection can undermine trust, especially in codebases with sensitive information, large monorepos, or strict token budgets, where developers want precise control over what gets sent to a third-party model.
This kind of feedback is emblematic of how Anthropic and its competitors are iterating on Claude Code based on real developer usage rather than purely theoretical design. Configurability—letting users toggle automatic file attachment on or off, or scope it to certain contexts—has become a recurring theme in developer tooling feedback loops. The broader trend here is the industry-wide recognition that agentic AI tools succeed not just on the strength of the underlying model but on the thoughtfulness of the surrounding product decisions: what gets included in context, how transparent that inclusion is to the user, and how much control developers retain over an increasingly autonomous coding assistant. As these tools proliferate, granular controls like this one are likely to become table stakes rather than optional refinements.
Read original article →