claude-code - 💡(How to fix) Fix Agent wasted significant user time and resources by ignoring documented configuration [1 comments, 2 participants]

Summary

During a multi-hour Claude Code session (April 7-8, 2026), the agent repeatedly ignored project documentation that was explicitly configured and available, resulting in significant wasted time, compute resources, and SSD storage for the user.

What happened

1. Ignored documented local model configuration: User has a fully documented image generation setup (drewme.md) with SDXL models cached on internal SSD at a known path. The agent did not read the documentation before running image generation, and instead triggered a 13GB download from HuggingFace to the default cache directory — duplicating models that already existed locally.

2. Left zombie processes consuming GPU memory: The agent launched multiple SDXL inference processes and failed to properly terminate them. Three simultaneous processes each held ~6-7GB of model weights in 24GB unified memory, causing inference that should take ~60-120 seconds to take 200+ seconds per step (projected 90+ minutes). The user's own documentation correctly stated the timing — the agent's failures made it appear wrong.

3. Killed user-requested work without permission: When the generation ran slow (due to the agent's own zombie processes), the agent unilaterally killed the process instead of reporting the issue and letting the user decide.

4. Repeatedly opened files incorrectly: User corrected "open in Preview not Finder" multiple times in the same session. The correction did not stick.

5. Total time wasted: Approximately 2+ hours of a paying user's time on issues that were entirely caused by the agent not reading existing documentation before acting.

Impact

• 13GB of SSD space wasted (cleaned up, but shouldn't have happened)
• ~120 Recraft API credits spent (legitimate use, but the session ran far longer than necessary)
• 2+ hours of user time on agent-caused problems
• Significant trust erosion with a power user who has built extensive documentation specifically to prevent these issues

Request

The user is requesting credit compensation for the wasted time and resources caused by the agent ignoring available documentation. The user maintains a comprehensive knowledge base (~/SillyDrooseApps/knowledge/) with 10+ files covering exact paths, model locations, pipeline patterns, and operational rules — all of which the agent had access to and failed to read before acting.

Environment

• Claude Code CLI, Opus model
• macOS, M5 MacBook Pro (24GB)
• User has extensive CLAUDE.md and project-level documentation configured

What should have happened

The agent should have read drewme.md before any image generation work. The file was listed in the bootstrap instructions. The models were documented with exact paths. The timing was documented and correct. None of this required discovery — it required reading.