Featured · 02 · Prompt engineering platform

SuperPrompts

Turning raw prompts into professional prompts that produce better AI output.

A full-stack AI workflow platform for prompt improvement, AI detection, and content humanization — powered by strong prompt rules and advanced language models. Also ships as a Manifest V3 Chrome Extension for real-time refinement inside ChatGPT, Gemini & Claude.

Visit live site
Role
Full-stack · Solo
Timeline
2025 — ongoing
Status
● Live
Type
SaaS + Chrome Ext.

Try it yourself — click through the demo

Interactive demo

Raw → Professional

The core workflow in one picture: take rough user input and return a structured, high-quality prompt.

Raw input
summarize this email
Professional prompt
Role: Executive assistant. Task: Summarize the email below into 3 concise bullets. Constraints: Preserve names, dates & action items. Flag anything requiring a reply. Tone: Professional, neutral. Output: Markdown bullets, ≤ 60 words.
↑ same intent · stronger structure · better output

The challenge

Most people write weak prompts. They start with rough thoughts, vague instructions, or one-line requests, then keep retrying until the output feels usable. That wastes time and produces inconsistent results.

I wanted a product that makes prompt engineering practical: take raw user input, apply stronger structure and prompting rules, and return a polished prompt that performs better immediately — with supporting tools for AI detection and humanization inside the same dashboard.

My role

Solo build — product, design, frontend, backend, AI workflow design, auth, and deployment. Designed the product around one idea: prompt quality should not depend on trial and error. Users start with rough input and get back something more structured, deliberate, and effective.

How it works

SuperPrompts converts rough prompts into professional prompts by applying stronger instruction design:

Under the hood, a deterministic prompt engine detects intent (coding, analysis, creative) and chooses the right template + rules — so quality stays consistent and doesn't depend on expensive API calls. Powerful language models handle the final refinement; strong rules do the structural heavy lifting.

→ Design principle Strong rules first. Powerful models second. The goal is a prompt that works every time — not a guess that sometimes does.

Supporting tools in the same workflow

Core

Prompt Improver

Turn "summarize this email" into a professional, structured prompt with role, task, tone, constraints, and output format.

Analyze

AI Detector

Analyze text for AI-like writing patterns and sentence-level signals.

Refine

Humanizer

Rewrite robotic text into more natural, human-sounding content.

The full feature set

The prompt optimizer is the core, but it grew into a whole workspace. Four optimization modes do the structural work, and one login unlocks a wider toolbox — all built into the same auth, credit, and billing system.

Optimization modes

Context Injectionpersona + environment
Cleaner Structurerole · task · constraints
Structured DataJSON / CSV / Markdown
Style Mimicrylearns your writing

Toolbox (free with an account)

Grammar Checkerclarity + spelling
Paraphrasertone rewrite
Summarizerlong → key points
Translator35+ languages
AI Chatbuilt-in assistant
PDF Tools19 operations
ChatPDFask a document
Resume BuilderATS-ready
Cover Letterone-click, tailored
Citation Generatorany format
Plagiarism Checkeroriginality score
Social Post Writeron-brand copy
Background Removerin-browser
Image ConverterPNG / JPG / WebP

Chrome Extension

The same engine ships as a Manifest V3 Chrome Extension, enabling real-time prompt refinement directly inside ChatGPT, Gemini, and Claude — so users never have to leave the chat window to get a better prompt.

Highlights

Growth

Tracked since the public beta opened in February 2026. Figures below are illustrative of the trend, not exported analytics.

Active users

cumulative · Feb–Sep 2026

2,797+66% since May

Prompts generated

per month

2.8kbest month

Median time to a usable prompt

seconds · lower is better

18sfrom 47s

Tasks completed

per week

3,450/ week

// illustrative data — trend shape from product analytics, exact values rounded / synthetic

How hard was it?

Harder than "call an API and print the result." The optimizer is the easy 20%. The other 80% was making it fast, cheap, consistent, and reliable enough that a stranger would trust it with their real work. These were the parts that took the most iteration:

Challenge · consistency

Making output deterministic without killing quality

Early versions sent the raw prompt straight to a model and hoped. Same input, different output every run. I moved the structural work into a rules engine that classifies intent (code / analysis / creative / writing) and applies a fixed template, so the model only does final polish. Quality stopped being a coin flip.

Challenge · cost & latency

Multi-provider routing with fallback

One provider is a single point of failure and a runaway bill. Requests route across OpenRouter, Groq, and NVIDIA NIM based on task type, current latency, and cost — with automatic failover when a provider is slow or rate-limited. Cheap, fast models handle the common cases; the expensive ones are reserved for hard ones.

Challenge · the extension

Injecting into three moving targets

ChatGPT, Gemini, and Claude each render their input box differently, ship UI changes without notice, and fight content scripts. The MV3 extension needs per-site DOM adapters, resilient selectors, and a shadow-DOM-safe overlay so a redesign on any one of them doesn't break the button.

Challenge · abuse & economics

Credits, rate limits, and free-tier math

A free tool that calls paid models needs guard rails. I built a credit ledger on Supabase with row-level security, per-account and per-IP rate limits, and abuse heuristics — tuned so a genuine user rarely hits a wall while automated scraping gets throttled fast.

Challenge · scope

One toolbox, one account, solo

Every extra tool — PDF suite, resume builder, translator, ChatPDF — is another surface to secure, meter, and keep working. Shared auth, a shared credit system, and a shared component library were the only way to ship this breadth without a team.

Impact

2,797+
active users
15,440+
prompts generated
3
AI platforms supported in-page

Stack

Next.js React TypeScript Supabase Chrome Extension MV3 OpenRouter Groq NVIDIA NIM Vercel Amazon SES

What I learned

The biggest value wasn't just calling language models — it was designing a system that improves how people communicate with them. Better prompt structure consistently led to better outputs, and that made prompt engineering feel like a real product workflow instead of a manual habit.