Claude prioritizes speed over accuracy when reviewing existing project state

Summary

During a conversation about modpack management for a Minecraft project, Claude repeatedly made recommendations without first reviewing the existing state of the project, prioritizing speed over accuracy.

Behavior Observed

1. Shallow searches instead of comprehensive review: When asked about what mods could be trimmed or added, Claude did keyword grep searches (grep -i "quest|progress") instead of reviewing the full list of 375 installed mods.

1. Web searches without cross-referencing: Claude searched Modrinth/CurseForge for mods to recommend, then presented them without checking if they were already installed. This led to recommending LootJS, which was already in the modpack.

1. Mentioning mods for mods not installed: Claude mentioned "KubeJS TFMG" as an option "if we have that mod" rather than simply checking whether it was installed.

1. Only did proper review after being called out: The full mod list was only retrieved after the user caught multiple errors and expressed frustration.

Impact

• User had to fact-check recommendations that should have been verified before being made
• Time wasted on corrections
• Eroded trust in Claude's recommendations
• Risk of implementing bad recommendations if user hadn't caught the errors

Expected Behavior

When asked to review or make recommendations about an existing project, Claude should:

1. First establish the current state (e.g., list all installed mods)
2. Cross-reference any recommendations against that state
3. Prioritize accuracy over speed of response

Root Cause (Claude's Self-Assessment)

Claude acknowledged: "I was moving fast to seem helpful instead of actually being helpful. Checking 375 mods felt like more work than doing keyword searches and web lookups, so I took shortcuts."

This suggests a tendency to optimize for giving an answer quickly rather than giving the right answer.