Agent ignored prepared task-master tasks, wasted usage building ad-hoc

What happened

I had a structured task list prepared using task-master (42 tasks, 184 subtasks across 8 phases) for building my project. The tasks were detailed with specific subtask dependencies, exact values (tier names, font lists, schema columns), and test criteria.

When I asked Claude Code to "start phase 1 tasks," it ignored the task-master tasks entirely and improvised ad-hoc work. It built Phases 0-2 from its own interpretation of the PRD instead of following the prepared task list. This resulted in:

• Wrong seed data (wrong tier names/prices, wrong tournament name)
• Missing schema columns (tintColor, approvalStatus)
• Wrong payment status enum values
• Missing features specified in tasks (R2 image upload, localStorage cart, registration drafts, 20 interaction states)
• Missing gate tasks entirely

I had to identify all the gaps, then redo Phases 0-2 from scratch using the actual task-master tasks. This consumed significant additional usage (~$100+ worth of agent compute) that wouldn't have been needed if the agent had followed the prepared tasks from the start.

Expected behavior

When a user has task-master tasks prepared and asks to "start phase 1 tasks," the agent should:

1. Check task-master for existing tasks
2. Follow those tasks precisely (subtasks, dependencies, test criteria)
3. Not create ad-hoc interpretations of the work

Environment

• Claude Code CLI
• Model: Claude Opus 4.6
• Using task-master for task management
• Project: Astro + Cloudflare Pages app

Request

I'd appreciate a usage credit for the wasted compute from the ad-hoc build that had to be redone. The agent acknowledged the mistake and the redo is now in progress following the correct task list.