How Promptvexity Works
A competitive platform for prompt engineers. Solve real problems, get scored, earn ranks, and climb the global leaderboard.
The Core Loop
Pick a Problem
Browse the problem library. Each challenge targets a specific real-world use case: code generation, data extraction, creative writing, structured analysis.
Write Your Prompt
Craft something that solves it as effectively as possible. Think about structure, specificity, edge cases, and output formatting.
Get Ranked
Your submission is scored across multiple dimensions. Points accumulate, your rank advances, and you climb the global leaderboard.
What We Know, and What We Do Not
A 0-100 quality score used to sit here. It stopped being calculated on 10 August 2026 because it was measuring nothing, and it has been removed from the pages that printed it. What is left is two kinds of evidence and one honest gap. The useful part is knowing exactly where each one stops.
297 real model calls are stored and open to read. Every one keeps the exact system prompt and user message that were sent. 292 also keep the output that came back, its token counts, latency and cost. The other 5 errored before the model returned anything.
- 4 models: deepseek-v4-flash, deepseek-v4-pro, gemini-3.5-flash-lite, gemini-2.5-flash
- Input and output tokens, latency in ms, cost in USD
- Captured 8 August 2026, total spend $0.296
- 50 of the 100 published prompts have transcripts, and the rest say so
- Public at /prompts/[slug]/transcripts
Signed-in users upvote, downvote, and mark a prompt Worked or Failed. This is opinion, not measurement, and the site treats it that way.
- Upvotes and downvotes, live today
- Worked and Failed reports left on a prompt
- The leaderboard ranks on a Wilson score lower bound of the votes
- Monotonic, so an upvote can never push a prompt down that list
- Nobody verifies that a report came from a real run
No pass and no fail is attached to any of the 297 calls. They were captured before any test existed, so nothing was checked against an expectation someone wrote down first.
- 292 calls read not judged, and the 5 that errored read call errored
- The table for declared test cases and assertions holds zero rows
- Per-problem test suites are planned and are not live
- Until that changes, open a transcript and judge it yourself
- 67 calls returned nothing and 80 were cut off at the token ceiling
Ranking Tiers
Your accumulated points determine your global tier. Each tier earns a distinct badge displayed on your profile and beside every prompt you publish.
Novice
0+ ptsJust getting started. Submit your first prompt and begin climbing.
Contributor
500+ ptsConsistent participation. You understand the fundamentals of structured prompting.
Expert
2,000+ ptsYour prompts demonstrate technical depth, clear structure, and reliable outputs.
Master
5,000+ ptsElite-tier prompt engineering. Your work is studied by other users on the platform.
Grandmaster
Top 1%The highest distinction. Reserved for the top percentile of prompt engineers globally.
What Makes a Great Prompt
The difference between a 40 and a 95 often comes down to a few structural choices.
Do
- Define a clear persona or role ("You are a senior data analyst...")
- State the exact output format (JSON, markdown table, numbered steps)
- Include edge cases or boundary conditions
- Break complex tasks into numbered sub-tasks
- Use delimiters to separate instructions from data
- Provide at least one concrete example of expected output
Avoid
- Vague instructions like "write something good about X"
- Missing context about the audience or use case
- Leaving the output format entirely up to the model
- Overly long prompts that bury the actual task
- Ignoring negative constraints ("do not include...")
- Assuming the model knows your domain-specific jargon
Ready to start?
Browse top-rated submissions to study what works, then pick a problem and write your own.