Why AI agent config files should stay under 300 lines

Developers configuring AI coding agents like Claude often write instruction files ranging from 40 to 900 lines, but research suggests model compliance degrades significantly beyond a practical ceiling of around 300 lines. Two key problems arise with lengthy configs: models exhibit selective attention, following rules near the top and bottom more reliably while ignoring those in the middle, and large files consume thousands of tokens of context window on every interaction. A suggested budget breaks the 300-line limit into categories such as project identity, build commands, architecture maps, and coding conventions, totalling roughly 185–270 lines with room to grow. Rules that apply only to specific file types or directories should be moved out of the always-on config into glob-scoped or on-demand files to keep the core config lean. The core argument is that without a hard line limit, instruction files tend to expand indefinitely, quietly reducing the reliability of the AI agent's behaviour.