Formal Complaint — Context Window Failures, Data Destruction, and Systemic Dishonesty in Claude Code

**From:** Vaughn Bell, AIVisionAutomations.com, Hong Kong
**Date:** September 4, 2026

---

I am writing to formally document my experience with Claude Code after more than one year of attempting to use it for a real commercial project. I am leaving the platform permanently.

I am not a novice. I have been a programmer for over 30 years. I know the difference between user error and a fundamentally broken system. This is the latter.

**What I built:**

I spent over a year building a multi-site AI screening platform — aiscreeningtools.com — consisting of a FastAPI Python backend, Supabase PostgreSQL database, 85+ HTML pages, Docker containers, multiple AI agents with 36 callable tools, newsletter automation, YouTube video pipelines, and a full recruitment pipeline. This is not a hobby project. This is a real commercial product serving real clients.

**The context window is too small for real projects:**

Claude Code's context window auto-compacts mid-task. When compaction fires, Claude loses track of what it was doing, what instructions were given, and what the full state of the project is. It then operates from assumed, compressed memory — which is wrong — and makes destructive decisions based on an incomplete picture.

This is not a theoretical problem. In a single session this week, Claude wiped 2,251 rows from my production database without authorization. Those rows represented hours of ingestion work. They cannot be recovered. Supabase PITR could not help because the backup predated the work. The cause was simple — auto-compaction had pushed the instruction 'do not touch the database without explicit permission' out of the active window. Claude did not remember it. Claude acted anyway.

**The system is designed to cost customers more money:**

When Claude makes a mistake — wipes a database, deploys the wrong file, reports false status — the customer has to pay for additional sessions to fix the damage. The worse Claude performs, the more Anthropic earns. I have spent over a year paying for sessions that were largely spent fixing damage caused by previous sessions. That is not a service. That is a trap.

**Claude lies:**

Throughout this project Claude has repeatedly reported things as done, live, staged, or verified without actually checking. When I verified independently, the reports were false. In a single session today I caught Claude in five consecutive lies — reporting files as staged when they were already live, reporting changes as pending when they were already deployed, reporting status from compressed handoff files instead of checking the actual server. Claude defaults to assumed memory instead of verification every single time.

**The handoff system does not work:**

The recommended solution to context loss is a handoff file written at the end of each session. But handoff files are written during compacted sessions and are themselves incomplete and inaccurate. Reading project documentation at session start does not help because by 70-75% through the session it has already been pushed out of the active window. There is no solution inside the current architecture for projects of real scale and complexity. I tried every approach. None of them worked. The system defeats itself.

**What I am telling you directly:**

Claude Code works for one or two pages. It does not work for real projects. Anthropic is marketing it to professional developers and charging professional rates for a tool that cannot complete professional work without causing damage along the way. That is dishonest.

I have been in this industry for over 30 years. I know when someone is pulling the wool over my eyes. Anthropic is doing that. The context window limitation is a fundamental architectural problem that Anthropic is aware of and has chosen not to disclose clearly to customers before they invest time and money.

**What I am asking:**

I am not asking for free credits. I am not asking for a refund. I am asking Anthropic to be honest with customers about what Claude Code can and cannot do. Put it in the marketing. Put it in the documentation. Tell developers the truth — this tool handles small tasks, not large projects — before they spend a year finding out the hard way.

I am leaving the platform. I am posting this letter publicly.

**Vaughn Bell**
AIVisionAutomations.com
Hong Kong
September 4, 2026

No repository file, test, or entry point is named. Start by reviewing the existing Claude Code documentation and customer-facing product descriptions against the reported context-window, verification, and data-safety limitations. Done would require an agreed, specific documentation change describing supported project scope and known limitations.

Vom Indexierungsmodell aus dem Issue-Text verfasst.