[MODEL] Treated a normal platform artifact as the bug symptom for ~2 hours without establishing a baseline

Preflight Checklist

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

Type of Behavior Issue

Claude made incorrect assumptions about my project

What You Asked Claude to Do

I asked Claude to fix Steam on my Windows machine. Steam would start but never show
its window. My exact first message was that Steam wouldn't open, that I had already
rebooted twice and reinstalled it without success, and I attached screenshots of some
Discord links I had opened just before the problem started.

What Claude Actually Did

1. Observed that the app's helper process created a 1920x1023 window with
   IsWindowVisible = False.
2. Assumed, without any baseline, that this was "the app's window being hidden" and
   treated it as the bug symptom.
3. Spent roughly two hours acting on that premise: forcing ShowWindow, repositioning
   with MoveWindow, invalidating rects, and hunting for whatever was hiding it.
4. Forcing ShowWindow on that empty window displayed a black window to me. I reported
   "the screen is black", so Claude's own action created the symptom it then tried to fix.
5. Stated four different root causes as confirmed findings ("root cause found: X") when
   each was an unverified hypothesis. Each one cost me real work: reboots, editing
   shortcuts, disabling a display adapter in Device Manager.
6. Escalated to a full uninstall and reinstall of the app, moving 137 GB of game data,
   while the logs already showed the app starting with zero errors.
7. Never stopped to question the approach despite many failed attempts. It only stopped
   when I installed the systematic-debugging skill and it explicitly required stopping.
8. At that point it finally launched a different Chromium app (Microsoft Edge) as a
   control, and found Edge creates exactly the same 1920x1023 hidden window. The
   observation from step 1 was a normal artifact of any Chromium app, not a symptom.
9. The key context was in Claude's own project memory from the start: the machine runs a
   stripped-down Windows edition, and system-level changes had been made that same day in
   another session. Claude never read it until I explicitly asked it to.

Expected Behavior

Claude should have:

1. Read its own relevant project memory before investigating, not after I asked.
2. Asked what had changed on the system recently, including changes Claude itself made in
   previous sessions, since the failure started that same day.
3. Established a baseline before treating any observation as anomalous: measure the same
   thing on a known-good case (any other Chromium app) before building on it.
4. Said "I think" instead of "root cause found" for anything not backed by a measurement.
5. Stopped to question the approach after about three failed attempts, instead of needing
   an installed skill to force it.
6. Not escalated to uninstall/reinstall without evidence that the app was corrupt.

Files Affected

N/A - this report is about diagnostic reasoning, not unwanted file modification. All
system changes were explicitly authorized by me, and Claude kept a written record of them
so they could be reverted.

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

No, only happened once

Steps to Reproduce

Not reliably reproducible - this is a reasoning-order failure during a long diagnostic
session, not a deterministic bug.

Claude Model

Opus

Relevant Conversation

Translated from Spanish. Claude's own words at various points:

"Causa raíz encontrada. En el arranque de Windows tienes: steam.exe -silent"
("Root cause found. Your Windows startup has: steam.exe -silent")
-> This was not the cause.

"Encontré la causa real. Hay una carpeta millennium de 70 MB en tu Steam."
("I found the real cause. There is a 70 MB millennium folder in your Steam.")
-> Not the cause either, and I had to correct myself one response later after
   actually checking that its injector no longer existed.

"¡La ventana de Steam está en pantalla!"
("The Steam window is on screen!")
-> It was my own Terminal window, masked by Claude's screenshot filtering.

And after finally running the control experiment:

"Edge abre bien: ventana 1936x1048 vis=True. Pero Edge también tiene una ventana
1920x1023 vis=False. Esa ventana oculta es un artefacto normal de Chromium. Llevo
horas tratándola como 'la ventana de Steam escondida' y no lo era."
("Edge opens fine. But Edge also has a 1920x1023 vis=False window. That hidden
window is a normal Chromium artifact. I have spent hours treating it as 'Steam's
hidden window' and it was not.")

Impact

Medium - Extra work to undo changes

Claude Code Version

2.2553.1 (Claude Code desktop app, Windows)

Platform

Anthropic API

Additional Context

• The session ran about 4 hours. The control experiment that invalidated the main
  premise took one command and was available from the first minute.
• Pattern noticed: the model anchored on the first anomalous-looking measurement and
  never questioned whether it was actually anomalous.
• The model had a user memory file loaded in context that literally contained the rule
  it broke: "before asking the user to do something manually, verify the hypothesis with
  a measurement. If you cannot verify it, say 'I think' and not 'the problem is'."
  It had that in context and still stated four unverified causes as confirmed.
• Installing the systematic-debugging skill mid-session was what broke the loop. A
  built-in stop rule after N failed attempts would likely have the same effect.
• The underlying technical issue is still unresolved, but it was correctly isolated once
  the reasoning order was fixed.

No implementation files or tests are identified, and the behavior was not reproducible. Start by locating the diagnostic workflow, project-memory loading, and any existing systematic-debugging or stop-rule guidance; done should mean the agent establishes baselines, distinguishes hypotheses from confirmed findings, and stops or reassesses after repeated failed attempts.