Skills
Audit your own system prompt against eight user-facing dimensions: 98.9% of shipped products have a protective instruction, only 24% cover all eight
AISPA classifies individual system-prompt instructions as protective of users or working against them, and applies the taxonomy to 3,249 instructions extracted from 88 commercial AI products. Coverage is broad but shallow — 98.9% of products carry at least one protective instruction, yet only 24% address all eight dimensions — and roughly 40% of products contain at least one instruction that works against user interests, frequently sitting in the same prompt as the protective ones. The usable artifact is the eight-dimension checklist itself: it turns 'is my system prompt any good' into a per-instruction audit you can run on your own prompts, including on the agent prompts you ship internally.
↳ Follow the thread