Scoring Overview
Prism turns your telemetry into three personal scores — Speed, Skill, and Efficiency — plus a handful of supporting signals. This section covers the rubrics in detail.
How scoring runs
Section titled “How scoring runs”- You code with Claude Code (Prism plugin active).
- Telemetry (OTLP) and prompt text flow to the Ingest service.
- The Prism Engine batches telemetry to S3 and runs per-prompt scoring via an LLM rubric (
prompt_efficiency+ intent). - Layer 2 workers aggregate per-prompt scores into sub-session metrics (SSE, IE, CRR, FC, QR).
- Layer 3 assembles the five inputs into
Skill, plusSpeedandEfficiency, and persists them. - You view results via
/prism:score,/prism:report, or the dashboard.
The inputs
Section titled “The inputs”Skill is the central composite. Its five inputs, in order of weight:
| Input | Weight | Read more |
|---|---|---|
| Sub-Session Efficiency (SSE) | 45% | SSE |
| Prompt Efficiency (PES) | 20% | PES |
| Iteration Efficiency (IE) | 15% | IE |
| Context Reset Rate (CRR) | 10% | CRR |
| Flow Continuity (FC) | 10% | FC |
Speed and Efficiency have their own pages under Concepts → PRISM Scores. Speed is hours/week with a quality discount; Efficiency is SSE rolled up across the date range.
LLM-based scoring
Section titled “LLM-based scoring”The per-prompt rubric runs as an LLM call through the Optra gateway, not as in-process regex. Real-time feedback inside Claude Code is handled separately by the plugin’s real-time coaching.
Coaching notes
Section titled “Coaching notes”Every scored sub-session includes coaching notes — specific, actionable tips for the weakest input. They appear in:
/prism:scorecommand output- Dashboard PRISM Score pages
/prism:reportreview and/my/report- Session reviews on
/intelligence/reviews