Claude ignores its own established pattern mid-conversation, doubles token cost
In a single conversation, I had Claude build the same type of project folder for four clients in sequence. Each targets a VPS with SSH+mysql access. Same structure, same infrastructure, same pattern every time. An approved plan document specifying VPS infrastructure was in context throughout.
Claude built the first three correctly:
- Client A — SSH to VPS, direct mysql. Correct.
- Client B — Same pattern. Correct.
- Client C — Same pattern. Correct.
- Client D — Reverted to a deprecated shared hosting pattern (SFTP, no SSH, PHP script workarounds). Wrong.
For Client D, Claude generated 18 configuration files all targeting the wrong infrastructure. The plan document in context specified VPS for all client sites. The site had already been migrated to the VPS. When I told Claude to check, it found the VPS user, database, web server config, and SSL all in place. But by then it had already built everything against shared hosting.
I pointed out the mistake. Claude acknowledged it, then rewrote all 18 files from scratch with the correct pattern. Every token spent on the first pass was wasted.
The core issue: A plan document specifying VPS was in context. The pattern was established three times in the same conversation. Claude ignored both and silently used a deprecated approach. There was no ambiguity.
Token cost: 18 files built, thrown away, rebuilt. Roughly double the necessary cost for this task.
Expected behaviour: Follow the plan document in context and the pattern already established in the conversation. If uncertain, ask before building.
Actual behaviour: Ignored the plan, silently reverted to a deprecated approach, wasted tokens, required the user to catch the mistake and repeat instructions.
---
Environment: Claude Code (VSCode extension), claude-opus-4-6, macOS
This issue has 2 comments on GitHub. Read the full discussion on GitHub ↗