Righting Sentences User Guide

Last updated: January 2026

Production URL: https://righting-sentences-573737403792.us-east1.run.app/

GitHub Project: https://github.com/saxtonmd/righting-sentences

Current version: 3.0.4

Desktop user interface

Mobile User Interface

1. Purpose and Scope

Righting Sentences is a web-based chat application that generates better sentences from generative AI models. Prompts provide instructions to the models. The models then respond. Responses are expected to be more like standard written English. This means they should not contain repetitive triads, em dashes, or many words that genAI models frequently use. The goal is to make models that generate more human-like text.

How it works:

Users prompt an OpenAI, Anthropic, or Google model
A set of instructions encourages the model to return a response using its typical patterns
The program captures the model's response
- The first response is saved as a JSONL file in a Google bucket
- The first response is scored as an example of Bad' model behavior
A new prompt forces the models to generate a second response with editorial corrections
- The second response is saved as a JSONL file in a Google bucket
- The second response is scored as an example of Good' model behavior
The program then forces changes if the models violate the restrictions
The final output then appears
The Good and Bad examples will fine-tune a foundational model

1.1 Who this is for

Writers
Editors
Journalists
Marketers
Researchers
Assisstants
Analysts
Data enthusiasts
Software developers

1.2 Setting Realistic Expectations

This is a generative AI application. However, it is also a data science experiment. While Righting Sentences is like many public genAI apps, it is not the same.

Expect occasional bugs in the user interface.
Expect some responses to be slower to generate because of the two-prompt system.
Expect periodic odd sentence constructions, and expect to edit the text.

The application is under continuous development.

1.3 Warnings

Like other genAI apps, Righting Sentences can make mistakes. Furthermore, word pairings might be incorrect because the application determined that the correct method violated its instructions. While this behavior is by design, it should not be considered production-ready.

Do not rely on Righting Sentences as a guaranteed source of truth. Generative AI can make mistakes.
Do not use this software to replace advice from a doctor, lawyer, electrician, or other relevant professional.
Do not use this software to generate content straight to production.

2. Getting Started

2.1 Google Authentication

To use the application, you will need to sign in using the Google account associated with this project.

Navigate to https://righting-sentences-573737403792.us-east1.run.app/
Select Sign in with Google
Enter the Username and Password for the Google account associated with this project
Click Continue to accept the Terms of Service

You will be at the application's main screen.

3. Left Sidebar

The left sidebar contains:

Your chat history
The model selector
A menu for app functions

To access each collapsed section:

Hover the cursor over the Model selector to reveal the model selector and New Chat button
Hover the cursor over the Menu to reveal Account, Settings, Help, and Sign Out options

3.0.1 Open and Close the Sidebar

To close: select the X at the top right of the sidebar
To open: select the hamburger menu (☰) at the top left

3.1 Chat History and Controls

3.1.1 Chat history list

Each row in the chat history list shows:

The conversation title
The model used

To open a conversation:

Navigate to the left sidebar
Select the desired conversation

3.1.2 Delete a conversation

To delete a conversation:

Navigate to the chat history list
Select the trash icon to the right of the conversation

Notes

Deleting a conversation removes it from your chat history

4.2 Settings

4.2.2 Color

Righting Sentences offers nine theme colors. The default color is Blue. Other color selections are:

Purple
Pink
Red
Orange
Yellow
Green
Teal
Cyan

Selecting a new color changes the buttons, the user conversation bubble, and the outline of various page elements across the platform.

4.2.2 Custom Instructions

4.2.3 Model Instructions

4.3 Model Selector

The model selector controls which model is used for a new conversation.

4.3.1 Select a model

Navigate to the model selector in the left sidebar.
Select the model you want.
Select New Chat to start a conversation using that model.

4.3.2 Important: models cannot be changed mid-chat

Once a conversation has messages, the model selector is disabled for that conversation.

To change models:

Select New Chat.
Select a different model.
Continue in the new conversation.

4.3.3 Available models

The model list can change over time. Typical configured models include:

Provider	Example model names	Notes
OpenAI	gpt-4o; gpt-5.2-pro	Used for general writing assistance and editing.
Anthropic	claude-sonnet-4-5-20250929	Used for alternative writing style and reasoning.
Google	gemini-3-flash-preview	Used for fast drafting and variations.

4.3 Main Chat Window

The main chat window contains:

The header (conversation title and model indicator).
The message transcript.
The message input box and file upload controls.

4.3.1 Header

The header shows:

The conversation title (center/left)
The current model badge (top right)

4.3.2 Messages

Messages appear as bubbles:

Your messages are on the right.
Assistant messages are on the left.

Assistant messages can include formatting such as:

bullet lists
headings
code blocks

4.3.3 Copy assistant output

Each assistant message includes a copy control:

Navigate to the assistant message you want.
Hover the message bubble.
Select the copy control at the bottom-right of the bubble.

5. Writing and Editing Workflow

5.1 How to ask for what you want

GenAI performs best when you provide clear context and constraints.

Recommended prompt pattern:

What you are writing (email, report, press release, article, résumé).
Your audience (customer, internal team, general public).
Desired tone (formal, neutral, friendly, direct).
Length constraints (e.g., “under 150 words”).
Required content (facts, names, dates).

Example:

Rewrite the following paragraph for a general audience. Keep it under 120 words and use a neutral tone:
...

5.2 What responses should look like

The assistant should typically return:

A polished version of your text, and/or
Clearly labeled alternatives (Option A, Option B), and/or
A short explanation of major edits (when helpful).

If the assistant output is not what you want:

Select a specific change (tone, length, structure).
Ask for a second version rather than editing line-by-line.

5.3 Editorial behavior (two-pass system)

Righting Sentences uses an editorial process designed to reduce “AI-sounding” language.

What you may notice:

Responses may be rewritten to sound more natural and human.
Some words or phrasing patterns may be avoided by design.

6. File Uploads

You can upload files to a conversation so the assistant can reference them when responding.

6.1 Upload a file

Navigate to the message input area.
Select Upload.
Select one or more files.

After upload:

Files appear as “chips” above the input box.
If the file can be parsed, it may show a “Parsed” status.

6.2 Remove an uploaded file

Navigate to the file chip above the message box.
Select the remove (×) control on the chip.

Removing a file deletes it from the conversation’s context, so the assistant will no longer reference it.

6.3 Supported file types

The application is designed to accept:

Text: .txt, .md, .csv, .tsv, .json
Documents: .pdf, .doc, .docx
Presentations: .ppt, .pptx
Spreadsheets: .xls, .xlsx
Images: .png, .jpg/.jpeg, .gif, .webp, .bmp, .tiff

Notes:

Parsing quality varies by file type and content.
Very large or image-heavy documents may be partially parsed.

6.4 File size limits

The default maximum upload size is 25 MB per file (configurable by the administrator).

6.5 How uploaded files are used by the model

If a file is successfully parsed, the extracted text is attached to the conversation context so the model can reference it when answering your questions.

Tip:

Ask questions that cite the file: “In the uploaded document, what are the main recommendations?”
If the assistant says it cannot see the file, re-upload a text-friendly version or paste key sections into the chat.

7. Settings

Navigate to Settings from the left sidebar menu.

7.1 Theme

Select a theme:

Light
Dark
System (matches your OS preference)

7.2 Primary color

Select a primary color to change accent styling (borders, icons, buttons).

7.3 Custom instructions

Use Custom Instructions to provide information about your preferences (tone, style, constraints).

7.4 LLM instructions

Use LLM Instructions to specify global rules for how the assistant should respond.

8. Account

Navigate to Account from the left sidebar menu.

The account page shows:

Your Google profile photo (when available)
Your name and email
A Sign Out button

9. Data Storage, Privacy, and Access

9.1 What is stored

The system stores:

Conversations and messages (including your prompts and assistant responses)
Uploaded files and file metadata
Optional research logs for the editorial system (when enabled)

Because this is a data science project, assume that your prompts, responses, and uploaded content may be reviewed by the project administrator for research and quality improvement.

9.2 Where data is stored

Data is stored in Google Cloud services:

Conversation metadata and message content are stored in Google Cloud Firestore.
Uploaded file objects are stored in Google Cloud Storage in a private bucket in the project’s Google Cloud environment.

9.3 Who can access data

Data access is restricted to:

The project owner and administrators for the underlying Google Cloud project (Saxton Publishing).

9.4 Third-party model providers

When you send a message, the conversation content may be transmitted to the selected model provider to generate a response.

When you upload a file:

Text may be extracted locally for common formats.
In some cases (especially images), file content may be sent to a model provider for extraction (OCR/parsing) depending on configuration.

10. Troubleshooting

10.1 I cannot sign in

Confirm you are using a Google account.
Confirm your account has been invited.
If the problem persists, contact Saxton Publishing.

10.2 I cannot change the model

This is expected once a conversation has messages.

To change models:

Select New Chat.
Select a different model.

10.3 My file upload fails

If you see an upload error:

The storage bucket may not be configured or may not exist.
The server may not have permission to write to storage.

Contact the administrator with the exact error message.

10.4 The assistant says it cannot access my file

Confirm the file chip is visible above the message box.
Upload a smaller version of the file, or a text-based export (for example, export a Word doc to .txt).
For scanned PDFs or images, upload a clearer image or provide the text in the chat.