Subtext
Prompt Intelligence for Power Users
Context & Role
PRODUCT
A Chrome extension + backend service for analyzing and improving AI prompts
MY ROLE
Solo founder / designer / full-stack developer
TIMELINE
2024–2025 (ongoing)
STACK
Chrome Extension, React, FastAPI, Firebase/Firestore, Redis, OpenAI API
Problem
As AI tools like ChatGPT became part of daily workflows, I noticed three recurring issues:
1. People don't know what makes a "good" prompt
Prompts are written by trial-and-error, with no systematic feedback.
2. High-value prompts get lost
Great prompts live in random chats, docs, or sticky notes. There's no memory, analytics, or versioning.
3. Teams can't standardize
Everyone rewrites the same instructions instead of sharing battle-tested prompt patterns and guidelines.
Core Question
How can we turn prompt-writing from a messy, individual habit into something trackable, improvable, and shareable?
Hypothesis
If we build a lightweight layer on top of existing AI chats that:
- Captures prompts in context
- Analyzes them with clear criteria
- Turns those into reusable, team-ready patterns
Then people will write better prompts faster, get more consistent outputs, and start building a shared library of "prompt playbooks."
Users & Use Cases
Primary Users
Students & Solo Builders
Using ChatGPT all day
Product Teams
Experimenting with AI workflows
Founders
Writing prompts for support, content, or coding agents
Example Scenarios
A founder wants to refine a system prompt for their support chatbot
A student wants to make sure their prompts sound less "AI-ish" and more human
A small team wants to capture and share the 10 prompts that work best for their product
Solution Overview
Subtext is a GitHub-style "lint + analytics" layer, but for prompts, embedded directly in ChatGPT via a Chrome extension.
How it Works
User clicks a floating action button (FAB) injected into the ChatGPT UI
Subtext sends the prompt to the backend for analysis
The backend returns:
- • A structured scorecard (clarity, constraints, persona, context, tone, risks)
- • Rewrite suggestions and alternative phrasings
- • A quick guide explaining why the prompt is strong or weak
User can preview and apply improvements directly in the chat box
Over time, prompts and their "performance metadata" can be stored (via Firestore) and reused as templates.
Architecture & Implementation
Frontend (Chrome Extension + UI)
- • Injected into ChatGPT using a Shadow DOM host to avoid style conflicts
- • Floating action button opens panels for: "Analyze Prompt", "Preview Rewrites", "Guides & Patterns"
- • Uses XMLHttpRequest to communicate with backend API
- • Toast notifications for success/error states
- • Accessible labels & keyboard focus management
Backend (FastAPI 1.2.0)
Endpoints
- • Prompt analysis (single call)
- • Full analysis with OpenAI
- • Prompt clarification
- • Prompt guides
Firestore Collections
- • Users
- • Prompts
- • Events
- • Templates
- • Vectors
- • Redis: Caches repeated analyses to reduce cost and latency
- • OpenAI + embeddings: Evaluates prompts and stores vector embeddings for similarity search
- • Safety & robustness: Circuit breaker for OpenAI failures with fallback responses and retries
- • Security: Input validation & sanitization for all user data
Design Process
Discovery
I shadowed my own usage and friends' workflows:
- • Screenshotted messy prompts
- • Marked when a prompt "worked" vs "failed"
- • Identified patterns in strong prompts: clear role/persona, specific constraints, context and examples, explicit success criteria
Defining the "Scorecard"
I turned those patterns into a simple rubric. Each analysis returns ratings on these dimensions + explanation:
UX Constraints
Extension must feel lightweight, not like a second app
Never block the user from sending their prompt
Show value in under 3 seconds
Iteration
Early versions were too verbose: walls of text about how to improve.
I iterated to:
- → Short bullet improvements
- → A single "Try this rewrite" button
- → Optional "deep dive" for users who want the why
Current Status & Impact
Status: MVP Fully Working in Development
Impact (Qualitative)
Users become more aware of:
How vague their prompts are
When they forgot to specify tone or format
When they're asking for too many things at once
Rewrites often read more natural and human, improving the chance of passing AI-detection in writing-heavy contexts.
Key Learnings
Prompt quality is teachable
Users quickly internalize patterns after a few analyses.
Inline > separate tool
People won't go to another tab just to "fix" a prompt. Embedding into ChatGPT was crucial.
Explain the "why," not just the "what"
The educational layer (small guides) turned out as important as the raw score.
Local-first and privacy matter
Next steps include moving more analysis on-device or via local LLMs for users who care about privacy.
What's Next
Short-term Roadmap
Template Library
Save successful prompts and organize them by use case (coding, research, content, teaching).
Team Features
Shared prompt libraries, usage analytics, and tags for org-wide "prompt standards."
Local Models
Optional integration with local LLMs (e.g., via Ollama) for offline / private analysis.
Multi-tool Expansion
Extend beyond ChatGPT to other AI tools (Notion AI, Gemini, etc.), making Subtext a general "prompt intelligence layer" for any AI interface.