Fantasy football draft assistant: one-shot prompt challenge

Prize: $100. Entries close Friday, September 4, 2026 at noon Eastern. The winner is announced Saturday morning, September 5.

Enter the challenge here

Write one prompt that works as a live fantasy football draft assistant. Paste it into Claude, ChatGPT, or Gemini before your draft, and it should guide you through all 15 rounds while you type in what is happening. The best one wins $100 and gets published here and in the Prompt Teardown newsletter.

What you are building

You are writing a single prompt, not an app. There is no code to ship and no account to connect.

A working entry does four things. It gives you a full draft plan the moment you paste it in. It keeps track of who has already been taken. It tells you who to pick when you are on the clock. And it does all of that fast enough to read while a 90-second timer is running.

What AI draft assistants got wrong last season

These are real complaints posted by people who tried using AI during their 2025 drafts:

• It suggested players who were already taken, and it got worse as the draft went on.
• It gave injury news from the wrong season.
• It wrote three long paragraphs while a 90-second clock was running.
• By round 11, it had not noticed that every running back on the team had the same bye week.
• It mixed up 2025 rosters with 2026 rosters.
• It repeated the same rankings you can find for free anywhere.

The last one matters most. A prompt that just prints consensus rankings is not worth pasting in. The prompts that win here will be the ones that reason about your roster and the players still on the board.

The league every entry is tested in

Every entry is tested in the same league, so the scores can be compared against each other:

Every finalist is tested in the same FantasyPros Draft Wizard mock draft simulator, at the league settings above. This challenge is not affiliated with FantasyPros.

You can practice at the same settings at draftwizard.fantasypros.com/football/mock-draft-simulator/settings/. Drafting is free, and a free account is needed to see your score at the end. Set League Type to 2026 Season, Scoring to Half PPR, Draft Type to Snake, 12 teams, and Opponent Pick Logic to Advanced. The roster is QB 1, RB 2, WR 2, TE 1, FLEX 1, K 1, DST 1, Bench 6. The simulator defaults to WR 3 and FLEX 0, so both of those need changing.

Every finalist is drafted from the same middle draft position, between pick 5 and pick 8. The exact position is not published in advance so nobody can build for one specific slot, and it is announced with the results.

Where the player data comes from

The model does not reliably know who is on which team in 2026. If your prompt leans on the model’s own memory of players, it will invent things, which is why “mixed up 2025 rosters with 2026 rosters” is on the list above.

Two approaches are allowed. You can paste your own rankings into the prompt, though a top-100 list will eat most of your word budget. Or you can write the prompt to work from what the user types during the draft, which costs no words and matches how a real draft goes.

The second is closer to real use and harder to write well, which is exactly what is being scored.

Rules

Read these carefully. If your entry breaks one of them, it cannot be judged.

• You must be 18 or older to enter.
• One entry per person. If you submit more than once, only your first submission is scored and the rest are ignored. Entering under multiple accounts or names disqualifies all of your entries.
• Your prompt must be 1,000 words or fewer. Counted in Google Docs, on the exact text you paste into the entry form, once, at submission. Rankings, instructions, examples, and placeholders all count. A placeholder counts as the words inside its brackets: [MY DRAFT POSITION] counts as three words, [POSITION] counts as one, and naming it shorter is allowed. The values judges fill in for testing do not count, and neither does anything typed during the draft. Going over means disqualification, so check before you submit.
• Placeholders are allowed. Your prompt can include placeholders like [MY DRAFT POSITION] that the user fills in before sending. Judges fill in the league facts listed on this page and the draft position. A placeholder for anything else, like a team name or a keeper list, gets deleted before testing, so it earns you nothing. Deleted or not, its words were already counted at submission.
• One paste, one plan. Judges paste your prompt as the first message, alone, nothing added. Its first reply must be a complete draft plan. A prompt whose first reply asks questions instead of giving a plan is not scored. After the first reply, the only thing judges type is what happens in the draft.
• Your prompt must work on a $20 plan: Claude Pro, ChatGPT Plus, or Google AI Pro. Testing uses the default model on each one, with no extra reasoning mode turned on.
• Judges test in a brand-new chat with web search off, memory off, and custom instructions off. The prompt gets no help from anything outside itself.
• Your prompt must not need a paid data subscription.
• You must have a public GitHub repository containing the prompt and a license file.
• The prompt you paste into the entry form is the one that gets tested. Editing your repository later does not change your entry or what gets linked.
• Do not copy another entry. If two entries are very similar, the earlier independent creation wins, and commit history counts as evidence.
• An entry that is substantially identical to a publicly available prompt can be disqualified.

How to enter

Entry takes one form and closes Friday, September 4, 2026 at noon Eastern.

Enter the challenge here

You need five things, and the entry cannot be judged if any of them is missing:

1. Your Reddit username.
2. Your full prompt, pasted in.
3. A link to your public GitHub repository.
4. One sample output from your prompt, pasted in.
5. One screenshot showing your prompt working.

The sample output and the screenshot should come from the middle of a draft, around round 8 or later. The first few picks are easy for any prompt. The middle of the draft is where most of them fail, so that is what is worth showing.

The form also asks which model you built on (your finalist test runs on that model), a short name for your prompt, and your email. You are not signed up for anything when you give your email. It is used only for messages about this challenge.

How judging works

Judging happens in two steps, because a full mock draft takes about an hour and there is not enough time to run one for every entry.

Step one is screening. Every entry gets its screenshot read and its prompt pasted in for a quick test. The quick test checks that the first reply is complete, readable at a glance, and invents nothing. A screenshot can be staged, so it only decides whether an entry goes further. Screening is at the organizer’s discretion and is not appealable.

Step two is the full test. The best 5 entries get the full test. Each finalist is tested on the model they named on the entry form, playing a complete draft in the simulator described above. The simulator’s free tier allows five mock drafts a day, which is why five entries get the full test. During the test judges type each player’s name as picked, adding the team or position only when two players share a name. Nothing else gets typed, so no entry gets extra help. The pick clock is turned off during testing, so every prompt gets the same amount of time to answer.

Only the Draft Board tab is used during testing. It shows which players have been taken and nothing more. The simulator’s Suggestions tab and Cheat Sheets tab give FantasyPros’ recommendations, so they stay closed. The Rankings panel on the left stays open even on the Draft Board, so judges collapse it with the sidebar arrow before the first pick. The header still shows the simulator’s own top suggestion, and it is ignored. If those recommendations were followed, every entry would produce a similar team and the test would measure FantasyPros instead of the prompt.

If the prompt recommends a player who is already taken, judges reply “already taken” and draft its next recommendation, and the mistake counts against output quality. If the prompt offers several players at once, judges take the first one listed.

Drafting in the simulator is free. Seeing the draft score at the end requires a free FantasyPros account, so all five finalists are judged from the same free account.

Finalists are scored on three things:

Team quality is measured by the simulator’s draft score, the 0 to 100 number it reports on the results screen at the end of the draft. The same number is used for every finalist.

If two entries score the same, the shorter prompt wins. The prize is awarded to the best valid entry regardless of how many entries arrive. Judging decisions are final.

What the winner gets, and the roundup

The winner receives $100, paid by Amazon gift card or PayPal, whichever the winner prefers, plus a full write-up on this page and a feature in the Prompt Teardown newsletter.

You do not have to win to get published. Every entry that passes screening and agrees to publication is listed in a roundup here on Saturday, September 5, in time for anyone still drafting that weekend. You choose on the entry form whether to be included, and whether to be credited by name or listed anonymously.

Every repository in the roundup is linked at the commit that matches the submitted entry. That protects entrants two ways. Nobody can improve their repository after the deadline and claim a better result, and nobody’s linked work can drift away from what was actually tested.

Who runs this challenge

This challenge is run by the Prompt Teardown newsletter.

This challenge is not sponsored by, endorsed by, or associated with Reddit. By entering, you release Reddit from any liability related to this challenge.

Sponsors

Sponsors fund the prize pool. A sponsor receives their name and link at the top of this page, a mention in the results write-up and the newsletter, and a “prize sponsored by” credit in the winner announcement.

A sponsor receives no judging role, no entrant data, no repository access, and no entrant contact information. Sponsorship money received after entries close rolls to the next edition.

To sponsor, email [email protected]

Official rules

• You must be 18 or older. Void where prohibited.
• Entry is free. Winners are chosen on merit by judging, not by chance or by public vote.
• The prize is paid by Amazon gift card or PayPal, whichever the winner prefers, where payment is legally possible.
• The winner must confirm by direct message from the Reddit account used to enter, before payment is sent.
• If the winner does not claim within 7 days of the announcement, the prize goes to the runner-up.
• The prize amount can increase before entries close, and locks at the moment entries close. No other rule changes after the announcement.
• Entrants keep all rights to their prompts. Publication happens only if selected on the entry form.
• Accepting the prize includes publication of the winning prompt and repository, with attribution per your form choice. A winner who declines publication passes the prize to the runner-up.
• Entrant data is never shared with sponsors.
• Entry data is deleted within 60 days of the announcement, except published material.
• Moderators, judges, and members of their households cannot enter.
• Judging decisions are final. Screening is at the organizer’s discretion and is not appealable.
• An entry may be disqualified for breaking these rules, or removed if it turns out to be someone else’s work.
• Entrants release Reddit from any liability related to this challenge. This challenge is not sponsored by, endorsed by, or associated with Reddit.
• Questions: [email protected]

Results

Results are published here Saturday, September 5, 2026. This section will hold the winning prompt, the roundup of every published entry with a repository link, the draft position used in testing, and notes on what separated the top entries from the rest.

If you are reading this before Saturday, the challenge is still open.