Product

Status: Draft. This document defines what SkillNet is, who it’s for, and what it does.

The current company-first product remains the implemented baseline. The future audience model for organization and individual deployments is defined in audience-modes.md.

What is SkillNet

SkillNet is a learning system that builds the right training experience for each person, from the knowledge that already exists in their company.

Not a course catalog. Not a static LMS with AI bolted on. A system that reads what a company knows — manuals, procedures, protocols — and turns it into training that adapts to who is learning, not just what is being taught.

Open source, self-hosted, one instance per company. Not multi-tenant — by design.

It doesn’t compete with enterprise offerings. It exists for the companies that those offerings don’t serve.

The core idea: the same company knowledge should produce different training experiences for different people. Not because an admin configured it that way — because the system understands each person’s role, level, and progress, and builds accordingly.

Roles

Role	What they do
Admin	Uploads documents, reviews generated content, assigns training, sees team progress
Employee	Learns, practices, asks questions. The experience adapts to their level and pace

Content types

Type	Purpose
Course	Modules + exercises + evaluation. Structured learning path, generated from company documents
Manual	Reference material. Employees consult when they need it. Organized for lookup, not learning
Chatbot	Per-content chatbot. Employees ask questions about the material and get answers grounded in it

Content generation

The primary way to create content:

From documents — Upload a PDF, manual, or protocol. A team of AI agents extracts themes, designs a structure, generates modules and exercises, reviews quality, and produces a course + manual. The admin reviews at two checkpoints before anything reaches employees.

The generation pipeline is a LangGraph state machine with 10 nodes, 7 specialized agents, and 2 mandatory human checkpoints. See content-generation.md.

Future generation methods (not in MVP):

From conversation — tell the AI what you know, it structures the course
From scratch — give it a topic and level, it generates original content
From living docs — when source documents change, affected courses are flagged for regeneration

Exercises

Multiple types, defined by the content itself. Examples include tests, practical cases, real-world tasks (“do this and tell me if it worked”), and others to be determined as the product evolves.

Every exercise includes an explanation citing the source material. Answers are evaluated either deterministically (test, true/false, fill_blank) or by an LLM with a rubric (practical_case, dialogue).

Tracking

Employees complete courses. The system records what they know how to do:

Exercise attempts with scores and timestamps
Skill levels that increase when exercises are passed
Spaced repetition scheduling for review
Deadlines and enrollment status

The admin sees team progress, skill gaps, and alerts. How exactly this is presented is open — the data model supports multiple views.

Adaptation

SkillNet adapts at two levels:

Level 1 — Content generation (offline, expensive): The course is generated once from company documents. But the generation process already considers the target audience: the admin specifies who the course is for, and agents adjust Bloom levels, exercise difficulty, and examples accordingly.

Level 2 — Experience adaptation (real-time, cheap): Each employee sees the same course differently based on their profile:

A beginner gets more theory lessons and guided examples
An experienced employee skips to exercises and gets harder practical cases
The tutor agent adjusts its explanations based on conversation history and past performance
Spaced repetition schedules exercises for review at the optimal moment

Level 3 — Adaptive regeneration (future, data-driven): After a course has been taken by enough employees, the system identifies patterns: which modules have low pass rates, which exercises are too easy or too hard, which topics generate the most tutor questions. This data feeds back into the generation pipeline to regenerate weak modules automatically.

Signal	What it tells us	Action
Low pass rate on a module	Content is unclear or too difficult	Regenerate module with simpler explanations
High tutor questions on a topic	Employees don’t understand from the course alone	Add examples or a dedicated lesson
Fast completion + high scores	Content is too easy	Increase exercise difficulty or add advanced module
Abandoned course at a specific point	Friction or disengagement	Investigate and adjust that section
Spaced repetition failures	Retention is poor	Adjust FSRS parameters or add reinforcement

How adaptation works in practice is open. The data model already captures all the signals needed (exercise_attempts with scores, timestamps, tutor chat logs, spaced_repetition table). No schema changes required — just the logic to act on the data.

Learning Loop

The system learns from every interaction:

Employee takes course
    |
    v
Exercise attempts recorded (score, time, answer)
    |
    v
Skill levels updated
    |
    v
Spaced repetition schedules next review
    |
    v
Tutor chat logs questions and confusions
    |
    v
Admin sees patterns: skill gaps, struggling employees, weak modules
    |
    v
(Future) System flags content for regeneration based on real data

This is not a feature of the MVP. It is the direction the product moves in. Every table in the data model already supports it — the learning loop is a design constraint, not an afterthought.

Living Content

Company documentation changes. Policies are updated, procedures are revised, new regulations appear. SkillNet treats source documents as living, not static:

When a document is re-uploaded, the system detects what changed
Affected courses and manuals are flagged for review
The admin decides whether to regenerate or keep the current version
Employees see a version indicator so they know if their training is current

This turns SkillNet from a “generate once” tool into a system that stays in sync with the company’s actual knowledge.