Inspiration
AI is becoming a tool that everyone uses, but nobody practices using it. Most people type vague requests and then get frustrated when they don't receive the exact output they wanted. We drew inspiration from a joke website, opusfived.dev, which challenged users to get AI to change the color of a button without changing anything else, and from daily games like Wordle.
What it does
Promptless is a daily prompt-engineering game. Every day, everyone gets the same task, like:
"Get the AI to write a sentence of at least 8 words without the letter 'e'."
"Get the AI to output only valid JSON with exactly these keys."
You chat with an AI and try to get it to produce a reply that passes, in as few prompts as possible.
- Fewest prompts wins. Time is the tiebreaker on a daily leaderboard.
- The AI is lazy on purpose. It does the bare minimum and takes the most literal reading of what you write, so vague prompts fail. You win by being clear and specific, which is the real skill.
- No cheating by typing the answer. If the answer appears in your own prompts, it doesn't count. You have to get the AI to produce it.
- One play per day. Like Wordle, each device gets one leaderboard score per day, plus a shareable result: "I solved today's Promptless in 2 prompts."
- 25 tasks across easy, medium, and hard, rotating automatically each day.
How we built it
| Layer | Details |
|---|---|
| Backend | Python with FastAPI, served by Uvicorn. Four simple API endpoints handle loading the task, sending prompts, saving scores, and loading the leaderboard. |
| AI | The Groq API running gpt-oss-20b, with a custom system prompt that gives the AI its lazy, literal personality. |
| Judging | We deliberately did not use an AI to grade the AI. Every task has a plain Python checker (regex, JSON parsing, word counts), so judging is instant, free, and gives the same answer every time. |
| Database | MongoDB Atlas stores the leaderboard, with an automatic in-memory fallback so the game keeps working if the database is unreachable. |
| Frontend | Plain HTML, CSS, and JavaScript. No framework and no build step. |
Team roles
| Member | Role |
|---|---|
| Johnny | Backend: FastAPI server, API endpoints, and Groq AI integration |
| Solomon | Frontend: HTML, CSS, and JavaScript game interface |
| Joshua | Data: MongoDB Atlas leaderboard and the in-memory fallback |
| Lilian | Game design: the 25 daily tasks and their win conditions |
AI-assisted development
As beginners, we used Claude (Anthropic's AI assistant) to do most of the heavy lifting on the code. We made the design decisions: the game concept, the tasks, the scoring rules, the anti-cheat rules, and how the app should behave. Claude helped turn those decisions into working code, explained each change so we understood what we were shipping, and helped us debug. We tested everything ourselves, and playing our own game is how we found most of the problems listed below.
Challenges we ran into
- Players could just type the answer. At first, every task could be solved by copying and pasting the task into the prompt. We built a no-copying rule into the website to prevent this.
- The AI was too smart. We built tasks around things AI generally struggles with, like counting words, but then used a reasoning model that solved them on the first try anyway. We fixed this by making the AI deliberately lazy so that it does the bare minimum and players have to earn their wins.
- Clocks. The win screen and the leaderboard calculated time separately, so users saw two different times on screen. We moved timing to the server so every screen shows the same number.
- Spam. Refreshing the page started a new game, so one person could spam our API and leaderboard. We fixed this by giving each browser an ID that can only be used once per day.
Accomplishments that we're proud of
- We went from an idea to a real, playable game as a team of beginners in one weekend.
- The game rewards a real skill, writing good prompts, instead of luck.
- Win conditions are deterministic Python, not judged by AI.
- We split the work so that three people could work in parallel without breaking production.
What we learned
- How to build and deploy a full-stack website.
What's next for Promptless
- User stats so players can track their progress over time.
- Richer tasks that go beyond text, where the AI actually does things, like changing a color on the website.
- Smarter anti-cheat so users can't type something close to the answer and get it on the first try.
Log in or sign up for Devpost to join the conversation.