3,847 words. That was the system prompt one developer had been using with Claude for eight months, packed with coding standards, workflow requirements, project context, personality preferences, and error-handling rules - everything, all at once.
It was also the root cause of every conversation going off the rails.
The discovery came while debugging why Claude kept losing the thread on longer tasks. The culprit wasn't the complexity of the requests or a limitation in the model. It was the instruction block itself. A prompt that long creates competing priorities for the model to balance, and the more constraints you pile on, the more likely Claude is to quietly violate one while trying to satisfy the others.
The fix was to break it apart. Instead of one monolithic system prompt, use multiple shorter, focused prompts for different contexts. A prompt for code review. A different one for scaffolding new files. Another for debugging. Each one narrow enough that there's no ambiguity about what matters most in that session.
This tracks with how Claude's context window - the amount of text the model holds in working memory at once, roughly equivalent to a long document - actually processes instructions. More context isn't always better. A long instruction block buries the signal in noise, and the model ends up making judgment calls about which parts matter most in any given exchange. That's where unpredictability creeps in.
The practical implication: if your Claude conversations keep drifting or producing outputs that feel almost right but miss something specific, your system prompt is worth investigating first. Start with the minimum viable instruction set and add constraints only when you hit a concrete failure. Developers who've tried this consistently report the same counterintuitive result - the shorter the prompt, the more reliable the output.
For Claude Code specifically, where context accumulates across a long coding session as files and edits pile up, keeping initial instructions tight matters even more as the conversation grows.