Core Concepts

Optra Prism is built on a few core ideas.

The problem

AI coding agents are powerful but opaque. Developers face:

• No prompt feedback — you don’t know if your prompts are efficient or wasteful
• Invisible throttling — rate limits silently slow you down with no visibility
• No cost visibility — token spend accumulates with no breakdown by session, model, or pattern
• No coaching — you repeat the same mistakes because nothing tells you what to improve
• No guardrails — no DLP, no budget caps, no access control for teams

Prism solves these by instrumenting the AI coding workflow and surfacing insights.

PRISM Score

The PRISM score is a composite quality metric for AI-assisted coding sessions. It measures five dimensions:

Dimension	Weight	What it measures
Prompt Quality (PQ)	25%	Specificity and decomposition of your prompts
Iteration Efficiency (IE)	20%	How quickly you converge and recover from errors
Verification Discipline (VD)	20%	Whether you review and validate AI output
Tool Use (TU)	10%	Selection and context of tool usage
Advanced Features (AF)	10%	Delegation to subagents and configuration of AI behavior

Each dimension has 2 metrics, scored 0–10. The composite PRISM score is a weighted average.

See PRISM Score for the full breakdown.

Telemetry pipeline

Data flows through four stages:

1. Capture — the plugin captures OTEL telemetry and prompt text during your session
2. Ingest — the ingest service (port 9005) receives OTLP data and publishes to NATS
3. Store — the engine’s S3 writer consumes from NATS and writes Parquet files to S3
4. Analyze — the engine scores sessions, detects patterns, and serves queries via DataFusion

Intelligence loop

Prism creates a feedback loop:

Code with AI  →  Capture telemetry  →  Score & analyze  →  Surface insights  →  Improve  →  repeat

• Real-time: the prompt advisor scores each prompt before submission
• Session-level: the engine scores completed sessions and detects waste patterns
• Trend-level: the dashboard shows improvement over days and weeks
• Recommendations: data-driven suggestions for model rightsizing, prompt patterns, and budgets