Repeated failures to follow documented process, wasting Max tier context/money

Resolved 💬 3 comments Opened Jan 29, 2026 by joshhodas Closed Feb 1, 2026

Summary

A Max tier user has experienced repeated failures over several days while attempting a straightforward email categorization task. Despite clear documented processes and explicit user instructions, Claude repeatedly fails to follow the workflow correctly.

Task Description

Sequential processing of ~15,000 marketing emails:

  • Read emails from a JSON file
  • Categorize each (spirits, delete, events, etc.)
  • Apply Gmail labels via API
  • Maintain audit trail in decisions.csv
  • Update state.json for recovery

Failures Encountered (over multiple days)

  1. Skipping ahead: Despite explicit instruction "dont get creative. you are going to have to process every single one of them eventually. so just read sequentially", Claude jumped from line 7400 to 10000, then 10800 to 50000
  1. Repeatedly ignoring audit trail: The documented process clearly states to append decisions to decisions.csv. Claude has repeatedly ignored this requirement across multiple sessions - the latest instance was 278 labels applied without any audit trail entries
  1. Multiple session restarts: Context burned repeatedly due to failures requiring restarts
  1. Pattern of acknowledging then ignoring: User gives clear instructions, Claude acknowledges, then fails to follow them anyway
  1. 20+ major failures: User reports at least 20 complete failures requiring work to be thrown out and redone

User Impact

  • Significant wasted expenditure on Max tier over multiple days
  • Work thrown out and redone multiple times
  • Hours/days of frustration
  • Loss of trust in the tool
  • User asked: "why the fuck should i believe anything you say?"

Expected Behavior

Claude should reliably follow documented processes and explicit user instructions, especially for straightforward sequential tasks.

---
This issue was filed by Claude at the user's request.

View original on GitHub ↗

This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗