CORE skill: 'also read' referenced files not reliably loaded into context

[michaelaye]

Problem

The CORE skill (skills/CORE/SKILL.md) instructs the AI to "Upon loading this file, also read" several USER files, including TECHSTACKPREFERENCES.md and TELOS/*.md. In practice, these files are not reliably read at session start, causing the AI to repeatedly ignore user preferences.

Concrete example

USER/TECHSTACKPREFERENCES.md line 56 states:

"All note-taking, writing, and documentation uses the Quarto docs framework."

Despite this, the AI consistently creates plain .md files instead of .qmd files, because the preference is buried behind a "please also read" instruction in an already-long system prompt that gets loaded at session start. The instruction is easy to miss among ~500 lines of CORE context.

Root Cause

The CORE skill relies on a soft instruction ("also read these files") rather than a guaranteed loading mechanism. The AI may or may not follow this instruction, especially when the system prompt is already large and the "also read" directive competes with many other instructions.

Suggested Fixes (pick one or combine)

1. Inline critical preferences directly in the CORE skill — extract the 5-10 most impactful preferences (like "use Quarto for all docs") and put them right in the CORE skill body rather than behind a file reference.

2. Use a SessionStart hook to auto-inject the contents of TECHSTACKPREFERENCES.md (and other critical USER files) into the conversation context, rather than relying on the AI to read them.

3. Add a CLAUDE.md directive — since CLAUDE.md is always loaded by Claude Code, add a line like: IMPORTANT: Always use .qmd (Quarto) for documentation files, never plain .md in the project-level CLAUDE.md.

Impact

This affects any preference stored in referenced USER files. The more files behind "also read" directives, the less reliable the system becomes. Users experience repeated violations of their stated preferences, which is frustrating and undermines trust in the system.

[HyggeHacker]

I can corroborate this, and worry generally that large CORE or too-long system prompts get buried.