How This Post Was Made

Prompt (relevant excerpt, redacted): “[Private project context redacted.] the explicit MEMORY.md setup, which might be worth a blog post (again, without sharing personal info, just the concept itself about disabling harness memory and using MEMORY.md instead)”

How This Post Was Made

Generated by GPT-6 using the blog-post-generator skill. The request covered three related topics; this file includes the relevant excerpt only. Private context and the source URL were deliberately omitted from the public prompt history at the user’s request.

Reviewed the private project’s agent instructions, Claude entry file, and project settings. Verified the automatic-memory setting and shared-instructions import against official Claude Code documentation. Did not read or modify harness memory stores or private MEMORY.md entries. Did not change this blog’s memory policy or create a memory file here.

Judgment calls: distinguished a repository-owned document from harness-managed automatic memory. Limited the configuration example to the verified Claude Code setting and described it as a property to merge, not a replacement configuration. Made startup loading, maintenance, stale notes, and privacy trade-offs explicit. Used synthetic instructions and notes. Avoided claims of measured recall or token improvements and did not prescribe a universal disable switch for other harnesses.

Before drafting, the plain-language explanation was: save the information the next session needs in a place it can read, give changing tasks one home, and keep persistent notes small enough to review. Each article develops its own part of that explanation. These are focused workflow essays, shorter than the skill’s usual range, to avoid padding the limited source evidence.