Product

Chrome extension

Version 1.0.0. The extension brings the PromptFlowEngine engine to the AI chat you already use. It reads the prompt you are writing, shows what is unclear, missing, redundant, or sensitive, and writes an improved version back into the chat box when you ask it to.

Not on the Chrome Web Store yet. Until it is listed, build it from a checkout of the PromptFlowEngine repository and load it unpacked (below).

What it does

Action	Result
Analyze	Quality score, token count, and findings grouped as sensitive data and security, ambiguity, missing context, instructions, expected output, and unnecessary content. Each finding says how to fix it.
Restructure	Rewrites a loose request into Objective, Requirements, Constraints, Edge cases, Testing, Acceptance criteria, and Expected output. Only sections the task needs appear, and added guidance is counted separately from your own lines. See how restructuring works.
Improve	Removes filler and repetition, then verifies that every sentence, requirement, code block, and variable survived.
Redact	Replaces API keys, passwords, tokens, connection strings, and personal data with placeholders such as `[REDACTED_API_KEY]`.

On supported sites a small launcher button opens the side panel, and turns red while your draft contains a secret. Results can be inserted into the chat box, copied, or moved back into the editor for another pass.

Supported sites

chatgpt.com
chat.openai.com
claude.ai
gemini.google.com
Any other page with a text box, after you click the toolbar icon on that page. You can also paste a prompt into the panel on any page.

Each site is described by a short list of selector hints, not site-specific code. When a site changes its markup, the extension falls back to finding the prompt box generically, so it keeps working instead of breaking.

Install

shell

# from the root of your PromptFlowEngine checkout
corepack enable && pnpm install
pnpm --filter @promptflow/browser-extension build

Open chrome://extensions and turn on Developer mode.
Choose Load unpacked and select apps/browser-extension/dist.
Pin PromptFlowEngine to the toolbar, then open ChatGPT, Claude, or Gemini.

Requires Chrome 116 or later. Other Chromium browsers that support Manifest V3 side panels also work.

How it is built

architecture

AI chat page
   │  content script: finds the prompt box, reads and writes it
   ▼
Service worker
   │  validates the request, picks the processor from your settings
   ▼
On-device engine (default)   or   PromptFlowEngine HTTP API
   │
   ▼
Popup / side panel: findings and the rewritten prompt

The on-device engine is the same @promptflow/engine code as the website, VS Code extension, MCP server, and API, bundled into the extension. Nothing is downloaded at run time.

Privacy and permissions

Permission	Why
`storage`	Saves your settings. No prompt text is ever stored.
`activeTab`	Reads the prompt box of the page you are on, only after you open the extension there.
`scripting`	Adds the prompt-box reader to a page outside the list above, on that same click.
`sidePanel`	Opens the panel beside the chat when you click the launcher button.

No host permissions, no remote code, no analytics, and no account.
It reads the prompt box only: not the conversation, cookies, or anything else on the page.
Prompt text is held in memory while the panel is open and is never written to storage.
The page itself cannot message the extension or read the panel.

Settings

Setting	Default	Effect
Where prompts are processed	On this device	Choose the HTTP API for exact token counts. See below.
API address	`https://promptflowengine.com`	Must be an https origin (or `http://localhost` for development).
Model for token counts	`gpt-4.1`	Any model from the catalogue.
Add engineering guidance	On	Off restructures using only your own words.
Launcher button	On	Shown on supported sites.
Warn about secrets while typing	On	Always checked on your device.

API mode

With the HTTP API selected, Analyze, Restructure, and Improve send the prompt to the API. There is no sign-in and no API key. Two rules still hold: a prompt in which the on-device scan finds credentials or personal data is not sent, and Redact always runs on your device.

If the API is unreachable, slow, or rate limited (60 requests a minute), the panel says which and offers a retry; switching back to on-device processing always works offline.

Limits

On-device token counts are estimates (characters ÷ 4), labeled as such. The API gives exact counts for OpenAI models.
The engine is rule-based. It does not call a model, so it will not invent missing requirements for you; it tells you what is missing.
Selector hints for each site are maintained by hand. If a site redesign hides its prompt box from the extension, paste the prompt into the panel.