Skip to main content
Question

Claude AI not finishing prompt requests lately

  • September 26, 2025
  • 1 reply
  • 324 views
Josh22

A recent update makes it so it lists out every excruciating detail on what it intends to do based on your prompt:

  • Let me check the exact spacing around that section:
  • Let me get the exact text:
  • Let me be more specific:
  • Let me check the current state of the file:
  • Let me try just updating the className:
  • Let me check the exact formatting:
  • Let me try the exact line number approach:
  • Let me try again with proper spacing:
  • Let me try a different approach - let me look at a larger context around this area:

50+ messages later and it stops processing and fails to complete the prompt. It doesn’t generate a version for restoring either. It just stops working. This only happens since it was instructed to let the user know every detail it’s doing.

1 reply

disdar
  • disdar
  • New Member
  • March 13, 2026

This is a known failure mode that got worse when Claude Code became more verbose about its reasoning steps. The "Let me check... Let me verify... Let me try again" loop isn't just annoying — it's actively consuming your context window. Every one of those narrated steps costs tokens, and on a complex file edit you can burn through the context budget on internal monologue before the actual work is done. By message 50+ you're hitting the ceiling and the session dies without a restore point because it never reached a clean completion state.

The practical fix is to add a directive to your CLAUDE.md telling it to minimize narration and act rather than announce — something like "Do not narrate intermediate steps. Execute directly and report only what changed." You can also try starting fresh sessions for discrete tasks rather than letting one session accumulate 50 messages of back-and-forth edits on the same file. For managing multiple Claude Code sessions without losing track of where each one is — Remocode sits alongside agents like Claude Code and Copilot in a split-pane terminal and lets you monitor output and respond via Telegram remotely, so at least when a session dies mid-task you catch it immediately rather than returning to a frozen screen. Anthropic is aware the verbosity increase has a real cost, so a config option to suppress step-by-step narration will likely come — but until then, CLAUDE.md is your best lever.

Terms & ConditionsCookie settingsAccessibility statement