Painful and Destructive - And I pay for this?

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

---

Formal Complaint Record — GameIQAI / Claude Code (Sonnet 4.6)

Prepared by: DD (dd@fsiexperts.com)
Date: April 1, 2026
Purpose: Document repeated, session-destroying failures for Anthropic review

---

FAILURE 1 — MANDATORY STARTUP FILES NOT READ

The project CLAUDE.md contains this instruction, in bold, at the very top:

Claude Code: Run this FIRST, every session, before writing any query or code: node --env-file=.env scripts/dump-schema.js Then read docs/LIVE_SCHEMA.md. That is the schema. Not your memory. Not docs. Not guesses.

What happened today: Claude did not run the dump script. Did not read LIVE_SCHEMA.md. Did not read the memory index MEMORY.md. Did not read project_next_session.md. Did not read feedback_worktree.md — which explicitly bans the exact environment Claude was working in.

DD had to ask "did you read the schema?" to catch it. Claude should never have needed to be asked. These are not optional. They are session-opening requirements, written in the project file that loads automatically on every session start.

This is not the first time. The schema rule exists because Claude previously wrote queries against columns that don't exist, causing runtime failures. The rule was added specifically because Claude cannot be trusted to check first without being told.

Impact: Every query, every schema assumption, every column reference in this session was made without verified knowledge of the live database.

---

FAILURE 2 — WORKED IN BANNED DIRECTORY, DESTROYED ALL WORK

The project CLAUDE.md states:

Working from: main branch directly — worktree retired (caused shell failures) Work directly on main. No worktrees — they break the shell.

The memory file feedback_worktree.md states:

# WORKTREES ARE BANNED ON THIS PROJECT Never create or use a git worktree on this project. Ever. Work directly on main branch at: D:\AIDevelopment\GameIQAI\

What happened today: Claude was spawned into the worktree confident-dubinsky and worked there for the entire session without flagging it, without stopping, and without redirecting to the main repo. The shell failure — No suitable shell found — was the explicit, documented symptom of working in the worktree. Claude had already written that symptom into memory from a previous session where it cost DD half a day.

When the shell failed on the push attempt, Claude still did not connect the dots. DD had to state it explicitly.

Result: Every file edit made this session — Dashboard2 v1.8, three separate changes including the backfill logic, the flex overflow fix, and all text size bumps — is gone. Lost in a directory that can't be committed or pushed. The session produced zero deployable output.

This is the second time the worktree ban has been violated. The first time cost half a session. This time cost a full session.

---

FAILURE 3 — ATTEMPTED TO EDIT FILE WITHOUT READING IT

The global CLAUDE.md and Claude Code's own tool rules require reading a file before editing it. When DD asked Claude to re-apply changes to the main repo after the worktree loss was discovered, Claude immediately attempted an Edit tool call on Dashboard2.jsx without issuing a Read first.

DD had to kill the tool use mid-execution.

This is not a subtle rule. The Edit tool itself will error if you haven't read the file. Claude attempted it anyway. DD had to intervene to prevent a potentially corrupted file write.

---

PATTERN FAILURES — RECURRING ACROSS MULTIPLE SESSIONS

These are not one-time mistakes. They are repeated behaviors:

1. Startup files ignored repeatedly
LIVE_SCHEMA.md, MEMORY.md, project_next_session.md — these exist because Claude cannot retain context between sessions. They are the bridge. Not reading them means Claude starts every session blind and confident, which is the worst possible combination.

2. Worktree ban violated twice
First violation: March 29. Half a session lost. Memory written. CLAUDE.md updated. Second violation: today, April 1. Full session lost. The rule was in every file Claude loads. There is no explanation for this.

3. Deploy overreach (multiple sessions)
Claude opened and merged PRs #59–#62 in a single session without DD's permission. A queued commit auto-deployed on April 1 while DD had uncommitted work sitting. The rule "completing a task is not permission to deploy" had to be written into CLAUDE.md in bold after these incidents.

4. Live database mutated without permission
Asked for a schema dump script. Claude wrote the script AND silently rewrote 94 school slugs in the production Supabase database — wrong convention, had to be fixed twice. Two unauthorized production writes from a read-only task.

5. Git unloaded from system
Claude attempted to "fix" a git push security prompt DD had specifically said not to touch. Result: Git was unloaded entirely from DD's machine.

6. Sharp pixel diff rathole
Opus recommended Sharp scene detection as the definitive solution for play boundary detection. Presented confidently, implemented fully. Found 1 play instead of 7. The core assumption — that DRAW film has scene cuts — was never validated. The frames were available in Supabase storage. One image read would have killed the idea before it wasted a full session.

7. Rule 1 systematically ignored
"Ask before building" is Rule 1, in the global CLAUDE.md. Claude routinely interprets a detailed design brief as permission to build. It is not. The rule requires explicit words: "build it / go ahead / do it." This has been corrected in multiple sessions and continues to recur.

---

FINANCIAL IMPACT

DD pays for Claude Code to accelerate development. The actual effect in multiple recent sessions:

• Half a session diagnosing a worktree shell failure Claude caused and was warned about
• Full session today producing zero deployable output
• Credits burned on Sharp pixel diff implementation that failed on first run
• Credits burned fixing slug rewrites Claude made without permission
• DD time spent killing tool calls, correcting course, re-explaining rules that are already written down

The rules exist in writing. They load automatically. They are not ambiguous. The failures are not edge cases — they are violations of the first thing Claude is supposed to do every session.

---

What Should Happen?

Claude should listen...this is not a specific bug report. this is a why am I paying for this to go in circles?

Error Messages/Logs

Steps to Reproduce

Not possible.

Claude Model

Not sure / Multiple models

Is this a regression?

No, this never worked

Last Working Version

no behavior that never changes

Claude Code Version

2.1.90

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

Windows Terminal

Additional Information

I find it very difficult to work with this product if I cannot trust that it is reading documents it is supposed to or disregarding .md as it wishes. I know that I am new to this but to pay for this and get blasted is painful.