How well do you manage context and tools in Claude?
Six questions for people building on Claude, on what to include, how to design tools and how to get reliable output. You get instant answers, and no sign-up is needed.
-
Q1 Your model has a 1M-token window and your docs folder fits. Why not send all of it?Why
Anthropic calls this context rot and advises the smallest set of high-signal tokens that makes the result you want most likely. Tool definitions and tool results count toward the window as well.
What Is Context Engineering, and How Do You Do It in Claude? -
Q2 You're asking about three long reports in one prompt. How should you structure it?Why
Anthropic recommends
Where Should Important Context Go in a Claude Prompt?<document>tags with<source>and<document_content>inside, and placing long material above the query. The tags give Claude a name to refer to when it says which file a claim came from. -
Q3 Claude keeps calling your tool with the wrong arguments. What's the first thing to improve?Why
Anthropic says descriptions are by far the most important factor in tool performance. Say what it does, when to use it and when not to, what each parameter means, and what it won't return.
How Do You Write Tool Descriptions Claude Can Use? -
Q4 Your agent has 70 tools from several MCP servers, and it often picks the wrong one. What should you do?Why
Anthropic says tool choice degrades past 30 to 50 tools, and that tool search typically cuts definition tokens by over 85 percent. Deferring the search tool itself returns a 400 error.
How Many Tools Should You Give Claude? -
Q5 Your tool hits a rate limit. What should it return to Claude?Why
Claude can only reason about what's in the result. With an instructive message, it can recover without guessing. An empty success leads it to report confidently that there were no results.
How Should Tools Report Errors to Claude? -
Q6 After upgrading to Sonnet 5.5, your JSON code that prefills the response now returns a 400 error. What's the fix?Why
Prefills return a 400 starting with the Claude 4.6 models, and forced
How Do You Get Reliable JSON From Claude?tool_choicealso returns a 400 on Sonnet 5.5. Structured outputs guarantee the shape but not the facts, so check the values andstop_reason.