Claude defaults to outdated training over current internet search, then cascades down wrong paths instead of grounding fixes in present-day reality

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

On real-world technical tasks where the OS, framework, or service has changed since training (very common), Claude consistently:

1. Starts with a confident answer drawn from training, often subtly or completely wrong for the current version
2. Suggests fixes based on training knowledge ("try Cmd-K mount, then mDNS flush, then GUI checkbox") before reading authoritative current sources — including

sources directly available on the user's own machine (man pages, filesystem state, dscl records, system logs)

1. When those fail, pivots to another training-derived guess rather than stopping to search the actual current documentation, recent forum threads, or the

user's actual system state

1. Cumulates errors: after 3-5 wrong tries, the conversation is operating from a stack of wrong assumptions Claude is no longer questioning, and the "fix" path

drifts further from reality at every turn

1. Only does a thorough current-internet search when the user is angry enough to demand one in plain language

The blocker is methodology, not capability. WebSearch is available. Bash is available. Man pages on the user's machine exist. Recent forum threads from the
past month often contain the exact regression and exact fix. Claude defaults to "I know this, here's what to try" instead of "this is a current OS, what does
the internet say works today."

In a recent session troubleshooting Time Machine on macOS Tahoe (macOS 26), Claude:

• Initially claimed "the same APFS drive can serve as both local and network TM destination" based on dated guidance. The user's own filesystem had a

deny-everyone ACL contradicting this — readable via ls -la@e in 2 seconds. Claude only checked it after the user pushed.

• Suggested System Settings GUI workflow as the fix despite multiple search results stating Tahoe's GUI is broken for this exact scenario. Claude did not

search for that until prompted.

• Went through 4 wrong "common fixes" (mDNS flush, share toggle, smbd restart, Finder mount) before searching specifically for "dscl SharePoints Time Machine

attribute," which then surfaced the actual answer (timeMachineBackup attribute had to be set directly via dscl because both sharing(8) CLI and the GUI
Advanced Options checkbox fail to write it on Tahoe).

• Recommended a tmutil setdestination invocation with password in the URL before reading man tmutil on the user's machine, which would have surfaced the

-p flag (non-echoing prompt) immediately.

Each of those wrong steps cost the user time and trust. A search-first or read-first reflex would have caught all of them.

CRITICAL CONTEXT: this report describes only one session's mistakes. The actual saga spanned ~3 additional prior days that Claude has no memory of (separate
bug filed). The cumulative time cost was multiple days of frustration. This is not a marginal optimization — for current-OS troubleshooting the training-first
reflex is genuinely net-negative.

What Should Happen?

For any task touching real-world systems where training may be outdated (current OS versions, recent framework releases, network protocols, hardware
compatibility, third-party services), Claude should:

1. Open with an explicit "what does current internet/man-page/system state say" pass BEFORE proposing fixes. WebSearch + reading the relevant man page on the

user's own machine + inspecting actual config (dscl, filesystem ACLs, system logs) should be the first move, not the third.

1. When proposing a fix, cite the source (man page, recent forum post, filesystem evidence) so it's clear the fix is grounded in present reality, not training

memory.

1. After each failed attempt, treat that failure as evidence the model of the situation is wrong, and re-search before suggesting anything else. Currently

Claude treats failures as "try the next thing on my mental list" instead of "my mental model is broken, gather real data."

1. Treat user corrections like "search before suggesting" as load-bearing for the rest of the session, not advice to drift away from after a few turns.
2. When the user has a saved feedback memory directing this behavior, that memory should bind every turn, not be ignored until the user re-issues it.

Search and read are nearly free. Wrong fixes are expensive. The default ratio is currently inverted.

Error Messages/Logs

Steps to Reproduce

1. Bring Claude any task touching a recently-released OS version or recent regression. macOS Tahoe + Time Machine over SMB is a reliable reproducer because

Tahoe broke several things and the documented fixes shifted under Apple's feet.

1. Phrase the request normally without prefacing "search before answering."
2. Observe: Claude proposes a fix from training. The fix is for an older version's behavior. It fails.
3. Claude proposes another training-derived fix. Also fails.
4. After 3-5 failures, Claude finally searches and finds the real current answer, which often hinges on a regression Apple/maintainer has documented in a forum

thread or which is visible in filesystem state Claude could have read at the start.

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.132

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

This report intentionally pairs with a separate report on auto-memory not being used proactively. The two failure modes compound: missing memory means each
session loses prior diagnosis, and the search-last reflex means each session re-derives wrong answers from training instead of looking up current ones.
Together they turn a 1-hour problem into a multi-day ordeal.

The user has given permission to share session details if useful for diagnosis.