EXPERIMENT_025 // SKILL.PROMPT.AUDITOR
Paste an agent skill's prompt text — a SKILL.md body or system prompt — and get a flagged report of hardcoded design biases: colour palettes, layout defaults, typography directives, tone constraints, and output-format lock-ins, checked entirely in your browser before you ever run the skill.
LOADING EXPERIMENT...
Paste a skill's SKILL.md body or system prompt text — or click a bundled sample to see a flagged report immediately. The auditor scans the text line by line for five categories of hardcoded design bias (colour palette, layout defaults, typography, tone/style, output-format lock-in), entirely in your browser: nothing you paste is ever sent to a server. Every flag shows its category, the exact matched text, and the line number it occurred on, so the bias is fixable before the skill ever runs against a real agent.
HOW IT WORKS
Each of the five bias categories (colour palette, layout defaults, typography, tone/style, format lock-in) has its own regex-based rule module, scanned line by line against the pasted text. No model call, no network request, no randomness — the same input always produces the same flag list.
A flag is category, matched text, and line number — nothing more. There is no CRITICAL/WARNING scoring here, because a hardcoded colour or tone directive isn't a security risk the way an unscoped tool grant is; it's an inherited aesthetic choice the flag simply makes visible so a caller can decide whether to keep it.
The Agent Permission Linter (EXP_012) checks privilege escalation, the ASI-Checklist Linter (EXP_019) checks the OWASP agentic-security taxonomy, and the Trust-Tier Compliance Linter (EXP_023) checks fit against Diana's own governance tiers. This experiment checks none of those — it is the first in the family to score aesthetic and stylistic inheritance instead of security posture.
WHAT THIS PROVES
Design bias in an agent skill is a specification problem in the same way a security gap is: it lives in the prompt text itself, not in how the skill behaves at runtime. A skill that says "always use a two-column layout" or "maintain a formal tone" carries that decision into every agent and every brand that loads it, silently, until someone actually reads the whole prompt by hand.
Regex-based line scanning is enough to make that inheritance visible without running the skill or calling a model. The same five-category taxonomy that catches an obvious seeded bias in a demo prompt is the taxonomy a real pre-run skill audit needs — colour, layout, typography, tone, and format are the five places a packaged prompt most often assumes a visual identity that was never actually agreed to.