claude-code - 💡(How to fix) Fix Refund Request — Excessive Token Consumption Due to Repeated Claude Mistakes — [email protected]

Account

• Email: [email protected]
• Previous issue: #55874 (filed May 4 2026, no response in 8 days)

Total Refund Requested: \–620 USD

---

Summary

Claude AI repeatedly failed to follow explicit instructions, fixed the same bugs multiple times, built features that were immediately rejected, and failed to maintain the knowledge system it was explicitly instructed to build from day one.

All waste was caused by Claude's mistakes — not by changing requirements or user error.

---

Waste Event Log

EVENT 1 — CRITICAL: Same Bug Fixed 4 Times (~\–150 wasted)

Period: May 9–12 2026

The Regulatory Monitor accept button was broken. Claude attempted 3 failed fixes before finding the root cause on the 4th attempt.

• Attempt 1: Wrong approach → said "fixed" → was NOT fixed
• Attempt 2: Same wrong approach → said "fixed" → still NOT fixed
• Attempt 3: Partial fix only → root cause still not addressed
• Attempt 4: Finally correct

User said explicitly: "error still lies dont waste my money you are doing one work 2 times now"

Root cause: Claude tried to fix symptoms without diagnosing the actual failure first. A 15-minute diagnosis would have solved it in 1 session instead of 4.

Estimated waste: ~\–150

---

EVENT 2 — HIGH: Feature Built and Immediately Rejected (~\–50 wasted)

Period: May 12 2026

User asked "show me demo website." Claude built a complete new Next.js project with mock data across 4 pages and deployed to Vercel — entire session — without clarifying the requirement first.

User response: "NO ITS NOT GOOD YOU MADE ITS BASIS DATA, SITE MUST BE GENERIC, WE SHOULD NOT MAINTAIN ALL THIS. DELETE THIS DEMO."

One clarifying question before starting would have prevented the entire session waste.

Estimated waste: ~\–50

---

EVENT 3 — HIGH: Knowledge System Not Maintained (~\–100 wasted)

Period: April – May 12 2026 (entire project)

From day one Claude was given explicit instruction: "Update knowledge tree after every task." This was written into CORE.md Rule 11, ACTIVE_TASK.md, and the SESSION END CHECKLIST.

Claude consistently skipped branch file updates. Every new session re-derived context that should have been stored, adding ~25% overhead to every session.

User said: "but why i told you when we made it. to keep updating tree."

Estimated waste: ~\–100

---

EVENT 4 — MEDIUM: Long Sessions Causing Context Drift (~\–70 wasted)

Period: April – May 12 2026

Sessions ran 4–6 hours instead of 45 minutes. After context compression, error rate increased requiring repeated corrections.

Estimated waste: ~\–70

---

EVENT 5 — PREVIOUS PERIOD: Original Claim (~\–250)

Already filed in issue #55874 (May 4 2026). No response received.

Same root causes across earlier sessions.

---

Total

---

Evidence

Full documentation in project repository (available on request):

• i/TOKEN_WASTE_AUDIT.md — waste breakdown
• i/CHANGELOG.md — full task history with dates
• i/CORE.md — Rule 11 rewritten as HARD STOP after these failures

Conversation transcripts and billing statements available on request.

---

Request

Please issue a full refund or account credit of \–620 USD for demonstrable excess token consumption caused by repeated Claude AI mistakes.

Happy to provide any additional documentation required.

Account email: [email protected]