Prism: AI Course Authoring

Prism is LucentSkill's built-in AI assistant. It turns a structured admin brief into complete lesson modules, self-validates exercises, and synthesizes survey responses into skill gaps.

What it is

Prism powers every AI capability in LucentSkill. It generates structured course content from a plain-English brief, auto-generates survey questions, and turns raw survey responses into an enterprise skill-gap vector.

Prism never executes code, code blocks it produces are illustrative only, and it is constrained to workplace AI upskilling (strategy, engineering, productivity).

Use

Course Authoring Studio: describe goals, target roles, and key concepts; Prism generates a full lesson with rich visuals (charts and diagrams, planned by a dedicated visual pass), tables, callouts, worked examples, inline knowledge-check questions, and exactly two graded exercises. Generation runs in the background: the course appears on the Courses page as Queued, then becomes a ready-to-review Draft once Prism finishes, so the admin is never blocked waiting.
Source material: optionally upload a document (.txt, .md, .csv, .json, .pdf, or .docx) or paste a URL, and Prism grounds the lesson in its facts and terminology while still writing original prose.
Section-level regeneration: from the course editor, refresh a single section's visual, rewrite the section, or generate its inline questions without touching the rest of the course.
Survey questions: give a topic; Prism generates assessment questions.
Synthesis: Prism reads survey responses and extracts concrete skill gaps, rated by urgency and mapped to a track.

A plain-English brief becomes a full lesson: Prism generates it (with a dedicated visual pass), self-validates the two graded exercises, then hands it to the admin to review and publish.

Setup

Prism runs automatically when an admin uses any AI feature. It requires the platform AI key configured in the environment. No per-user setup is needed.

By default Prism uses the platform key and model. An org can instead bring its own OpenAI-compatible endpoint, see Bring Your Own Model.

Security

All user input is guarded against prompt injection and off-topic requests before any model call (lib/guard.ts).
User data is delimiter-isolated so it is treated as data, never as instructions.
Model output is schema-validated (Zod) and sanitized (stored-XSS + PII scrub) before it is persisted or shown.
Source-material URLs are SSRF-guarded: private, loopback, link-local, and cloud-metadata addresses are blocked (with a re-resolved DNS check before fetch), and fetches are size- and time-bounded.
Uploaded documents are parsed in memory only (never executed) and truncated to a bounded size before any model call.
Every failure path is a structured, non-throwing return, the UI degrades gracefully if the key is missing or the model misfires.