audit-agent-context
Installation
SKILL.md
Audit Agent Context
Report first, edit only what the user accepts. Argue spend and headroom, which §1 prices, never adherence: a smaller context file has not been shown to make an agent follow instructions better, and the claim is uncheckable here.
1. Measure what loads before the first prompt
Work from git rev-parse --show-toplevel. Always-on is everything the model receives before
the user types:
| Source | How to find it |
|---|---|
| Root agent file | CLAUDE.md, AGENTS.md, and any other root file the harness reads |
Its @-import chain |
Every @path line, resolved transitively: an import pastes the whole file into every request |
| Output style | The one named by .claude/settings.json's outputStyle, if any |
| Skill frontmatter | name + description of every installed skill, deduped by name across skills directories |
| Memory index | The always-on index file, not the memories it points at |