What problem does it solve? Repository context files like AGENTS.md often fail silently: they exceed the 20KB truncation cap, get shadowed by higher-priority files, or trip the prompt-injection scanner, leaving the agent running with no project context at all. This Skill audits the file the way the Hermes runtime actually loads it, then grades its content quality. ## Core Features & Use Cases - Hermes loading checks (H1-H7): Detects 20KB truncation with 70/20 head/tail splitting, one-file-wins priority conflicts, injection-scanner tripwires that block the whole file, SOUL.md vs AGENTS.md misplacement, cache-stability issues, subdirectory nesting opportunities, and Codex 32KB block placement. - Content rubric scoring: Scores the Karpathy 12-rule completeness checklist plus per-line lints for missing-why prohibitions, 28-word run-on directives, vague terms, and unescaped absolutes, with a readability pass flagging sections an agent cannot parse in one read. - Structured scorecard output: Produces a prioritized report separating hard loading failures from content grades, with concrete fixes ordered by severity. - Use Case: A team has a 64KB AGENTS.md that the agent seems to ignore. The audit reveals the middle 10% is dropped by truncation and an injection-example line blocks loading entirely, then lists the exact sections to move into nested subdirectory files. ## Quick Start Audit the AGENTS.md file in this repository and show me the scorecard with any truncation or injection-blocking failures.
