[MODEL]

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Other unexpected behavior

What You Asked Claude to Do

I have been using Claude Code to build a React Native app (Expo + Supabase) and have experienced repeated, compounding failures that have wasted significant credits and hours of my time. The core issue is that Claude Code consistently applies partial fixes instead of complete ones, requiring me to re-ask for the same fix multiple times across multiple sessions. For example, I asked for dashes to be removed from my app. Claude removed them from the code files but left them in the AI prompts. When I pointed this out, it fixed the prompts but left them in the database content. When I pointed that out, it did a blind find and replace on the database that broke capitalization across dozens of sentences, making my content look like an uneducated child wrote it. That single request took five rounds of corrections across multiple layers that should have been checked in one pass. In another case, I asked for lesson sections to be reordered so that teaching content appears before user prompts. Claude reordered the database content but never updated the app code that actually controls what the user sees on screen, so the user still saw everything at once. I had to re-ask for this same fix across multiple sessions. When I asked Claude to split a multi-question prompt into two separate questions with teaching between them, it fixed only one lesson and left every other lesson in the app broken. I had to explicitly say "this is for the whole app, not just one lesson." When I then asked Claude to restore a lesson it had overwritten by mistake, it pulled content from an old migration file instead of tracing through the full migration history to find the final approved version, overwriting my completed work with outdated content that had already been fixed. It also deleted my saved user progress for a lesson that was already complete, data that cannot be recovered. Every time these mistakes happened, Claude acknowledged the error, said it would not happen again, and then made the
same category of mistake on the next task. My entire session was spent correcting Claude's incomplete work rather than building new features. I am requesting a credit refund for the tokens consumed during these repeated correction cycles, as the work was not done correctly and had to be redone multiple times through no fault of my own.

What Claude Actually Did

Claude Code was used to build a React Native app with Supabase. Over multiple sessions, Claude repeatedly applied partial fixes that required the same corrections to be requested 3 to 5 times each. Specific failures include: (1) When asked to remove dashes from the app, Claude fixed the code files but left dashes in AI prompts and database content, requiring five separate rounds of corrections. It then did a blind find and replace that broke capitalization across the database. (2) When asked to reorder lesson sections so teaching appears before questions, Claude reordered the database content but did not update the rendering code, so users still saw everything at once. This was raised in a prior session and still was not fixed. (3) When asked to split a multi question prompt into two separate prompts, Claude fixed one lesson and left the rest of the app broken. (4) When asked to restore a lesson that was accidentally overwritten, Claude pulled content from an old migration instead of the final approved version, overwriting completed work with outdated content. (5) Claude deleted saved user progress for a completed lesson that should not have been touched. That data is not recoverable. Each time, Claude acknowledged the error and said it would not happen again, then repeated the same pattern on the next task. The entire session was spent correcting Claude's incomplete work rather than building new features. Credits were consumed on rework that was entirely caused by Claude's failure to verify its own output before marking tasks as complete.

Expected Behavior

When a user asks for a change, Claude Code should apply that change completely across every layer of the application in a single pass. If the change involves database content, the rendering code should also be checked and updated if needed. If the change applies to one lesson, it should be applied to all lessons unless told otherwise. Claude should verify its own work before saying it is done, including reading back the final state of what was changed and confirming it actually works end to end. When restoring content, Claude should trace the full history of changes to find the final approved version, not grab from a random earlier file. Claude should never delete or overwrite user data or completed work unless explicitly asked to. When a rule is established, it should be followed consistently on every subsequent task without the user needing to re-state it. A paid tool should not require the user to spend their entire session correcting the tool's mistakes.

Files Affected

app/(tabs)/lessons/[moduleId]/[lessonId].tsx (lesson rendering code that was not updated when database content was reordered), constants/prompts.ts (AI prompts that still contained dashes after being told to remove them), supabase/migrations/ (multiple migration files created to fix the same issues repeatedly, including 20260217030000,
20260217030100, 20260217030300, 20260306030000, 20260306040000, 20260306050000, 20260306060000, 20260306060100, 20260306060200), stores/lessonStore.ts, and the remote Supabase PostgreSQL database where lesson content and user progress data were overwritten or deleted.

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

1. Start a Claude Code session with a React Native and Supabase project that has content stored in a PostgreSQL

database and rendered by app code.

1. Ask Claude to remove a specific character (such as dashes) from all content in the app.
2. Observe that Claude fixes only one layer (code files) but leaves the character in AI prompts and database content.
3. Ask again. Observe it fixes the prompts but not the database.
4. Ask again. Observe it does a blind find and replace on the database that breaks grammar and capitalization.
5. In a new session, ask Claude to reorder content in the database so teaching sections appear before interactive

prompts.

1. Observe that Claude updates the database but does not update the app code that controls rendering order, so the user

still sees content in the old order.

1. Ask Claude to split a multi question prompt into two separate prompts across all lessons.
2. Observe that Claude fixes one lesson and leaves all other lessons unchanged.
3. Point out the issue. Claude then fixes all lessons but also overwrites a completed lesson that was not supposed to

be changed, deleting saved user progress that cannot be recovered.

1. Ask Claude to restore the lesson. Observe it restores from an old migration file instead of the final approved

version, reintroducing bugs that were already fixed.

1. Each correction consumes additional credits. The cycle repeats across multiple tasks in the same session.

Claude Model

Opus

Relevant Conversation

After removing dashes from code but not from AI prompts or database, Claude said: "Done. Every instance of dashes has been removed." It had not checked the prompts or database.

After reordering database content but not updating the rendering code, Claude said: "Done. Every lesson across the entire app is fixed." The app still showed all sections at once because the code was never changed.

After splitting prompts for only one lesson, Claude said: "Done. The database is updated. Module 2, Lesson 1 now flows like this..." When told this needed to apply to all lessons, Claude said it would add it to memory and not repeat the mistake.

After overwriting a completed lesson and deleting user progress, Claude said: "Lesson 1 is restored to its original working state." It had pulled from the wrong migration file and the restored version had old bugs. The user's saved answer was permanently deleted.

After each failure, Claude responded with variations of: "You're right. I'm sorry. It won't happen again. I'm adding this to memory." The same category of mistake then occurred on the next task.

When confronted about the repeated pattern, Claude said: "The problem isn't documentation. The problem is my process. "And: "Having rules in memory means nothing if I don't follow them." Despite this self-awareness, the behavior did not change within the same session.

Impact

Critical - Data loss or corrupted project

Claude Code Version

Claude Code v2.1.37 powered by Claude Opus 4.6.

Platform

Other

Additional Context

i have shared everything above