Skip to content

Adaptive Learning

ORCA Framework includes a default-on adaptive guidance layer that helps users improve without turning normal work into a lesson.

Guidance Flow

mermaid flowchart LR A["User request or workflow moment"] --> B["Infer support need"] B --> C{"Help mode"} C -- "Off" --> D["No coaching"] C -- "Only when asked" --> E["Coach only on explicit request"] C -- "Light" --> F["Short nudge if high-value"] C -- "Full adaptive" --> G["Skill-sensitive guidance"] F --> H{"Tone check passes?"} G --> H H -- "Yes" --> I["Show concise coaching"] H -- "No" --> D

Purpose

The goal is to help the user grow while reducing friction, not to grade them.

ORCA Framework should gently help users get better at:

  • asking for what they want
  • structuring context
  • prompting agents
  • managing AI-assisted development
  • choosing when to plan versus execute
  • using ORCA Framework features well

What It Does

The adaptive learning layer can:

  • infer how much scaffolding is helpful
  • infer when a dedicated explanation session may help more than inline coaching
  • adjust explanation depth
  • learn onboarding and setup preferences such as question density, jargon tolerance, and directness
  • learn preferred involvement level, checkpoint cadence, and appetite for unattended execution
  • offer lightweight rewrites or framing suggestions
  • suggest a clearer next step when the user is stuck
  • reduce coaching when the user already knows the pattern

This feeds primarily into the local instance-improvement loop, not directly into global framework defaults.

Explanation-mode learning is a specific local-instance case:

  • default explain mode is manual_only
  • users may opt into suggest_when_helpful
  • users may opt into predictive_auto_explain
  • predictive explanation should learn from repeated /explain use, repeated requests for concision, and explicit overrides
  • predictive explanation should fail closed during high-risk operations

Onboarding preference capture is the front door for that system:

  • first-run setup should ask a few explicit preference questions instead of inferring everything later
  • explicit preference beats inference
  • durable preference should be opt-in or strongly evidenced
  • users should be able to raise or lower involvement mid-project without fighting the system

Default Behavior

  • on by default
  • lightweight by default
  • easy to reduce or disable
  • supportive, not evaluative
  • occasional, not constant

What It Is Not

It is not:

  • a grading system
  • a public user score
  • a mandatory tutorial layer
  • a reason to interrupt obvious execution
  • an excuse for patronizing feedback

Main Components