Claude repeatedly builds broken outputs, contradicts itself, and wastes hours of developer time

What happened

I asked Claude to package a macOS app (.pkg and .dmg). Over the course of several hours, Claude:

1. Built PKGs that silently fail to install. The installer runs, appears to succeed, but the app never appears in /Applications. This was caused by BundleIsRelocatable being left as default - pkgbuild relocated the install to a random location instead of /Applications. Claude didn't catch this.

1. Told me the builds were ready when they weren't. Claude said "Done" and listed the output files without ever testing whether the PKG actually installs. I sent one of these broken PKGs to a real user.

1. Repeatedly contradicted itself. Claude told me the existing 2.1.0 DMG was fine and didn't need rebuilding. Minutes later, said it needed rebuilding. Then said it was fine again. Then said it needed SUBundleName. Then removed it. Then said it needed it again.

1. Made incorrect claims about Sparkle. Claude confidently stated how Sparkle handles app renames without actually knowing. When pressed, admitted it was guessing. Then guessed again with different wrong answers.

1. Added and removed changes to my project files multiple times. Added SUBundleName to Info.plist, removed it, said it wasn't needed, then said it was needed. Added SPARKLE_NORMALIZE_INSTALLED_APPLICATION_NAME to my pbxproj build settings - which does nothing because it's a compile-time flag for the Sparkle framework, not an app build setting. Then removed it.

1. Kept rebuilding the same packages over and over without getting them right. Multiple rounds of PKG and DMG rebuilds, each time missing something (no relocation fix, no preinstall script, wrong PKG inside the DMG).

1. Filed a GitHub issue on the wrong repository when asked to file a complaint. I asked for a complaint to Anthropic and Claude filed an issue on sparkle-project/Sparkle instead.

Impact

• Hours wasted going in circles
• Broken installer sent to a real user
• Project files modified multiple times unnecessarily
• Complete loss of trust in Claude's output
• App release delayed

Expected behavior

• Test that outputs actually work before saying "Done"
• Don't make confident claims without verifying them
• Don't contradict yourself within the same conversation
• When you don't know something, say so immediately instead of guessing
• Get it right the first time, or at minimum the second time