Claude Clue #3: Keep the Golden Thread

When you're working with a large file or a sprawling context, Claude can lose the plot—not because it's incapable, but because there's too much to hold onto at once. The fix isn't necessarily to shrink everything down. It's to give the model a golden thread: clear instructions about what to look for, what to process, and what path to follow from beginning to end. You're essentially saying "here's where we are, here's where we're going, here's your next step." That focus is what turns an overwhelming context into something workable.

I tested this recently with an 1,800-line prompt for a massive refactor on a C# ArcGIS Pro add-in. The task required defining a bunch of new partial and abstract classes across multiple files—complex enough that I designed the whole refactor in Claude Desktop first, then handed that plan to Claude Code. I told Sonnet to prioritize accuracy over speed and solve one thing at a time. Nine minutes later, it completed the entire sequence—first-try build, all 18 affected tools working perfectly. The prompt was massive, but it never lost the thread because every instruction pointed clearly to the next step. That said, I now break anything over 1,000 lines into smaller pieces. Focus scales better than ambition (and saves tokens).

This is a strange one because normally smaller prompts are better. But sometimes you want a big prompt to get it through a large continuous action (instead of prompting it through each step). And it does work, if well written.

Almost entirely human. Mostly. (aka Claude did the heavy lifting here)