Claude Code repeatedly fails to follow explicit instructions to inspect and use provided source files, instead making assumptions and implementing changes based on its own interpretation.

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

I explicitly asked Claude Code to inspect the provided mockup/source files and use them as the source of truth before making changes to the NSPF app's UI.

In particular, I asked it to check the mockup files and the actual mockup source code to reproduce the layout, positioning, styling, card dimensions, camera angle, and other visual details accurately, rather than interpreting or approximating them.

What Claude Actually Did

Claude initially claimed to have checked the mockup, but implemented several elements based on its own interpretation rather than the actual mockup/source.
It repeatedly made changes that did not match the provided design, including incorrect pitch orientation, player/card positioning, card dimensions, and standings styling.
When I pointed out the discrepancies, Claude often made another attempt based on its interpretation instead of first inspecting the source files directly.
This happened repeatedly over several correction cycles, despite my explicit instructions to check the provided files.
Only after multiple corrections did Claude finally inspect the actual mockup source code and discover that some of its previous assumptions were wrong.
Claude itself eventually acknowledged that it had been "guessing for too long" and that it should have looked directly at the mockup code.

Expected Behavior

Claude should have:

Read the provided mockup/source files before implementing the requested changes.
Treated those files as the source of truth rather than relying on visual assumptions or approximations.
Compared the existing implementation against the actual source and identified the relevant differences.
Implemented the requested changes based on that information.
Only asked for clarification when the source itself was ambiguous.

Files Affected

Multiple NSPF project files related to the dashboard UI, player cards, pitch/lineup view, and standings components were modified during the correction cycles.

Permission Mode

Accept Edits was OFF (manual approval required)

Can You Reproduce This?

Sometimes (intermittent)

Steps to Reproduce

Start a Claude Code session in an existing project.
Provide Claude with mockup/design source files containing the exact UI implementation.
Explicitly instruct Claude to inspect those files and use them as the source of truth.
Ask Claude to implement or correct the corresponding UI.
Observe whether Claude actually inspects the provided source before making changes.
In the affected sessions, Claude instead makes assumptions about the design and implements an approximation.
Point out the discrepancies and explicitly instruct Claude again to inspect the source.
Repeat the process and observe whether several correction cycles occur before Claude actually reads the relevant source code.

Claude Model

Sonnet

Relevant Conversation

Claude initially said it had compared the mockup "only at a glance, not line by line."

After several correction cycles, Claude eventually said:

"I'll look directly at the mockup code instead of continuing to guess from screenshots."

Later, after inspecting the source, Claude acknowledged:

"I had been guessing for too long."

This is the key issue: I had explicitly instructed Claude to inspect the provided source files, but it repeatedly proceeded by inference instead, requiring multiple correction cycles before actually consulting the source.

Impact

Medium - Extra work to undo changes

Claude Code Version

Claude 2.9939.2 (d3e504) 2026-09-24T17:20:35.000Z

Platform

Anthropic API

Additional Context

I have noticed this type of behavior repeatedly during long Claude Code sessions, and it appears to become more frequent when approaching my weekly usage limit.

In this particular session, approximately 91% of my weekly usage had already been consumed.

I cannot establish that the usage limit caused the behavior, but the correlation has been noticeable enough that I would like to know whether approaching a usage limit can affect model routing, reasoning effort, context handling, file inspection, or instruction following.

The main issue is not that Claude made an incorrect implementation once. The issue is that it repeatedly ignored an explicit instruction to inspect the authoritative source files, made assumptions instead, and consumed additional usage through multiple unnecessary correction cycles.