はるか is the public professional name of a developer who writes about coding-agent behavior. Their detailed Opus 5 account focuses on instruction handling, file inspection, tool use and context loss.[1][2]
Start with Opus 5 is a really bad model →
Links
Ideas
The developer describes a session in which an agent treated a globally injected skill description as workspace content, skipped basic inspection and made edits based on an imagined directory. Other examples concern missed document sections, long-running task monitoring and user instructions lost across repeated compactions. The article connects these failures to interactions among the model, system prompt and harness. It is a personal troubleshooting account based on a few hours of use. Its broad benchmark accusations are the author’s opinion, not a controlled evaluation.[2]
Tools, writing & talks
Writing, research & talks
Start here
A developer’s critical account of instruction loss, inappropriate tool choices and context-compaction problems.
Read the concrete file-inspection and repeated-compaction examples. Coding-agent experience and access to your own tool logs help compare the reported failure modes with a reproducible session.
Author[2]Original source