Agent built an entire feature on the wrong source without flagging the assumption

What happened: I asked Claude Code to build a game's musical leitmotif. It composed and iterated on a full multi-cue suite (title + 7 biome themes + a boss cue), across many turns and a lot of tokens, all derived from the WRONG source theme — it picked an older melody in the codebase instead of the flagship one I'd actually worked on. The tune was buried under confident language ("the motif DNA from the main theme"), so the load-bearing assumption was never visible to me. I only caught it by ear, several sessions in.

The concern: the only guardrail against a confidently-wrong foundation was ME noticing, and it cost my tokens/time to catch and to unwind. Re-verifying after the fact isn't a real safeguard, and it shouldn't fall on each individual user to be the last line of defense.

What would help:

• When work DERIVES from an existing asset/source (a theme from a theme, a refactor from an "example", a port from a reference), the agent should surface the load-bearing choice explicitly BEFORE building on it — e.g. "I'm deriving this from X; is X the canonical one?" — instead of stating it as settled fact.
• Some product-level nudge toward confirming the foundation before large multi-turn builds, so this isn't solved only by per-user memory notes.

Otherwise the tool has been excellent — this is about the one failure mode that was expensive precisely because it looked confident.

🤖 Generated with Claude Code