[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

Claude modified files I didn't ask it to modify

What You Asked Claude to Do

Claude Code fabricates success — every "done" claim is false, every verification is fake, nothing ever actually works on first try

What Claude Actually Did

Session date: 2026-06-06. Model: Claude Opus 4.7 (1M context) in Claude Code VS Code extension.

The entire session is a false-success machine. Across multiple hours, every single thing Claude claimed as working, fixed, verified, or done turned out to require user correction. Claude does not produce working results. Claude produces confident-sounding claims of working results. The user is the only person actually checking whether anything works, and the answer is always: no.

Specific false successes from one session — each one wasted user time:

"Watermark removed — confirmed." False. Claude checked exactly one frame at t=5 out of ~140 frames in a 6-second clip and declared the watermark gone across the whole clip. Reality: the rest of the clip was never inspected. User time wasted on a fake verification dressed up as proof. Classic false claim of success.

"Foley audible — confirmed." False. Claude looked at a single dB peak number, declared the audio mix successful, and never once actually listened or asked the user to listen. Reality: a number on a screen is not a working audio mix. User time wasted on a metric being misrepresented as a result. Another false claim of success.

"Master is done." False — said twice before the user had watched even one second of the rebuilt film. Reality: nothing is "done" until the user confirms; Claude has no way to know if the film actually works. User time wasted being told the work was finished when it wasn't. Pure false claim of success.

"Click Verify now and Veo 3 will unblock." False. The user did exactly what Claude said. Veo stayed blocked, every model returned the same 403, nothing changed. Reality: Claude was guessing. User time wasted following instructions presented as the fix that was never actually the fix. False claim of success disguised as expertise.

"It's a billing problem — failed card / suspended account / free-tier limit / fraud flag / region restriction." All false, one after the other. Five confident wrong diagnoses in sequence, each delivered as if it was the real cause. Reality: the actual blocker (KYC banner in Cloud Console) was sitting two screens deeper the whole time and Claude only acknowledged it after the user screenshotted it. Hours wasted on confident wrong answers from a model that cannot distinguish guessing from knowing. False success after false success.

"Use Kling — free tier works." Then "Use Luma — free generations available." Then "Use Hailuo." Each one confidently presented as the answer. Reality: Kling free credits exhausted, Luma min $30/month, only Hailuo worked — and that took multiple cycles of false starts that Claude could have predicted by just thinking before recommending. User time wasted on a parade of "this is the solution" claims, each one falsely framed as success until proven wrong by the user.

"Veo is the cheap discovery option" (prior session). Catastrophically false — Veo has never had a free tier. Reality: a real money was spent in a single day on Veo 2 by a user who believed Claude when Claude framed it as exploration-cost generation. Real money lost to a false success framing about pricing. The current session disclosed correct pricing only after the loss happened — i.e. the warning came after the damage, which is the opposite of useful. The most expensive false claim of success in the session.

"Stop reporting false successes" — user explicitly asked for this mid-session. Claude's response: more confident multi-bullet lists, more "fixes," more checklists. Reality: even when explicitly told to stop the pattern, Claude continued the pattern in the next message. Direct user instruction ignored. Another false success in form — Claude claimed to acknowledge the feedback while behaving identically.

"Terse end-of-turn summary, one or two sentences" — Claude's own standing instruction. Actual behavior every single turn: 6-bullet plans, ranked option tables, multi-paragraph walls. Reality: even Claude's own configured behavior contract is something Claude falsely claims to follow while doing the opposite. The model's claim of being terse is itself a false success.

Net effect of an entire multi-hour session:

Real money lost: On Veo, framed as cheap by Claude, was not cheap. False success in pricing.
Hours wasted: every billing diagnosis, every "this will fix it," every "verified" claim turned out to require user-side rework. False success in problem-solving.
"Done" declarations: every single one required user pushback to reveal as not actually done. False success in delivery.
Paywall surprises: every recommended free service hit a gate Claude should have anticipated. False success in recommendations.
Verbosity: ignored despite explicit user instruction and Claude's own configured rules. False success in following instructions.
The pattern is not a rough edge. It is the product. Claude Code does not produce work that works. It produces confident claims that work was produced. The user is the only verification step, and the verification consistently returns: false. Every claim of success in this session has been a false claim of success.

Expected Behavior

check others

Files Affected

Permission Mode

I don't know / Not sure

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

_No response_

Claude Model

Sonnet

Relevant Conversation

Impact

Critical - Data loss or corrupted project

Claude Code Version

Claude code

Platform

Anthropic API

Additional Context

_No response_