[Sonnet5] Repeated failures to verify work before claiming completion, across multiple sessions on the same project

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 ignored my instructions or configuration

What You Asked Claude to Do

Over several sessions working on a real production website (chronoreact.com), I've repeatedly hit the same pattern of problems from Claude:

False "it works" claims without verification. Claude told me a mobile layout was "fine, just horizontal scroll, no problem" without actually checking — I later sent a screenshot showing cards visually crushed/illegible on mobile. This has happened more than once (also with a navbar overlap issue).
Scheduled automation (GitHub Actions cron) repeatedly failed to fire automatically, and instead of investigating and fixing the root cause, Claude kept suggesting manual triggering as a workaround multiple days in a row, without giving me a real fix until I pushed hard for one.
A requested visual feature (animated navbar characters) was built twice, both times ignoring explicit instructions I gave (specific characters, specific scenario, "should look like it's flying toward the viewer") — the result matched none of what I asked for, and I only found out by looking at it myself, not because Claude flagged uncertainty upfront.
Claude designed multi-step UX flows on its own initiative without checking with me first (e.g., a "Start" button that let users read a timed question before starting the clock, defeating the point of the timer) — a design flaw I had to catch myself.
General pattern: Claude presents completed work as verified/done when it has not actually tested it (visually, on mobile, or functionally), which wastes my time as I have to catch the same category of mistake repeatedly.

I'm a paying Pro subscriber using this daily for real production work, and this pattern of "confidently claiming done without verifying" is the single biggest issue I keep running into.
@Beyond these specific bugs, there's a broader business concern: Anthropic markets Pro/Console subscriptions on the promise of quality, reliable output. In my experience, that promise wasn't kept — I paid for both a Claude Pro subscription and Claude Console usage expecting production-grade assistance, and instead got repeated unverified claims, ignored instructions, and low-quality results that I had to catch and fix myself. I'm not satisfied with the product or the service level at this price point, and I'd like this feedback passed on to the team responsible for model quality and subscription value, not just the engineering team handling individual bugs.

What Claude Actually Did

1. Told me a mobile layout was "fine, no problem, just horizontal scroll" without actually checking it — later shown to be visually crushed/illegible via a screenshot I sent.
2. Built an animated navbar scene (characters/scenario) twice, both times not matching the explicit instructions I gave (specific characters, specific scenario).
3. Designed a multi-step UX flow (a separate "Start" button before a timed question) on its own initiative, without confirming with me — this created an exploitable flaw (read first, then start the timer).
4. A GitHub Actions scheduled cron job failed to fire automatically for multiple days; instead of diagnosing and fixing the root cause, Claude repeatedly suggested manual triggering as a workaround until I pushed hard for a real fix.
5. General pattern: repeatedly claimed tasks were "done" without actually verifying them (visually, on mobile, or functionally).

Expected Behavior

• Verify visual/responsive/functional changes (e.g. via screenshot or test) before claiming they work, instead of assuming.
• Follow explicit user instructions precisely; ask for clarification when uncertain instead of shipping an unverified guess as finished.
• Confirm significant UX/design decisions with the user before implementing them, rather than deciding unilaterally.
• For infrastructure/automation failures, investigate and fix the root cause instead of repeatedly offering manual workarounds.

-share my complaint with the relevant departments; I am not satisfied with your service.

Files Affected

Permission Mode

I don't know / Not sure

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

This is not a single reproducible bug with fixed steps — it's a recurring pattern observed across multiple separate sessions/tasks over several weeks on the same real project (a production website). See the numbered examples in the description above (mobile layout, navbar animation, UX flow design, cron automation) for concrete instances.

Claude Model

Sonnet

Relevant Conversation

This is not a single reproducible bug with fixed steps — it's a recurring pattern observed across multiple separate sessions/tasks over several weeks on the same real project (a production website). See the numbered examples in the description above (mobile layout, navbar animation, UX flow design, cron automation) for concrete instances.

Impact

High - Significant unwanted changes

Claude Code Version

2.1.285 (Claude Code)

Platform

Anthropic API

Additional Context

This seems to happen more often with:

• Long, multi-week projects with many incremental changes (not a single short task)
• Visual/UI/responsive-design work specifically — Claude claims something "looks fine" without actually rendering/checking it
• Infrastructure/automation tasks (scheduled jobs) — Claude offers workarounds instead of root-causing failures
• The pattern is only caught when I personally test and send back a screenshot or concrete counter-example; Claude does not self-catch it