Why We Built Measurably Privacy-First

Your AI prompts contain some of the most sensitive information in your workflow: proprietary code, business logic, customer data references, and strategic thinking. Any tool that wants to analyze your prompts needs to handle that data with extreme care.

We chose a local-first baseline: automatic measurement runs on your device and sync excludes prompt bodies. Explicitly saved content and explicitly invoked AI features have separate, disclosed boundaries.

How It Works

In the hardened extension developer preview, the intended per-prompt flow is:

1. Captures the prompt text in the browser tab 2. Scores it locally using our deterministic heuristic engine 3. Displays your score in a non-intrusive badge 4. Syncs only numeric scores and metadata to the cloud

Step 4 sends data like: { score: 72, dimensions: { specificity: 85, context: 68, ... }, tool: "claude", timestamp: "..." }

It does not send: your prompt text, any code you pasted, any error messages, or any response text.

Why Not Use AI to Score?

We could get more nuanced scores by sending prompts to an AI model for analysis. We deliberately chose not to for three reasons:

1. Privacy: Sending prompts to another AI service defeats the purpose 2. Responsiveness: Baseline scoring avoids a model-provider network request; actual runtime varies by device, browser, and prompt length 3. Cost: On-device scoring is free; API calls add up fast

Our heuristic engine isn't perfect — it can't understand semantic meaning the way an LLM can. But it evaluates useful structural and syntactic prompt signals without sending the prompt to an LLM for baseline scoring.

What Data Reaches the Cloud

Numeric dimension scores (6 numbers, each 0-100)

Overall AES score (1 number)

Which tool was used (e.g., "chatgpt", "claude")

Prompt length in characters and estimated tokens

Timestamp

In the hardened browser developer preview, only these derived fields are eligible for optional cloud sync after secure pairing is restored. Separate, user-invoked web AI and template workflows may process prompt text transiently as described in the main privacy policy.

Verification

The extension is available as a checksummed developer-preview archive loaded unpacked in Chromium browsers. The repository remains private and unlicensed; any future community source release is gated on an approved license, governance, and verified release artifacts.