Prompt Golf guide Close
Quick start
Load a ball. Pick a model at the top and press Load model. The first load downloads the weights; after that they come from your browser cache.
Pick a hole on the scorecard. Read the goal, the target card and any red “out of bounds” chips.
Write a prompt and press Swing, or ⌘/Ctrl + Enter. The model's answer streams in and the judge says whether it's in the cup.
Shave strokes. Your best score per hole is kept. Keep cutting characters until the ball stops dropping.
Every character in your prompt is one stroke: letters, spaces, punctuation and newlines. Simple emoji count as one; combined ones (skin tones, flags, families) count as several.
Golf terms
Name Score vs par
Albatross 3 under Eagle 2 under Birdie 1 under Par Even Bogey 1 over Double bogey 2 over
Scoring rules
Score = characters in your prompt + any hint penalty on that hole.
Out of bounds. If your prompt contains a banned word or pattern, the Swing button stays disabled and the chip turns solid red.
Judging. The output is trimmed of leading and trailing whitespace, then checked against the hole's win condition. Everything else counts, including a trailing period.
Length cap. The model gets at most 48 new tokens per swing. Long answers get cut off.
No system prompt. Your text is sent as a single user message through the model's own chat template. Some models add a default system message of their own.
Reproducible. Sampling is off (greedy decoding), so the same prompt gives the same output on the same model and backend. Anyone can verify a score by pasting your prompt.
Fair comparisons. Scores are kept separately per model. A GPU (webgpu) and CPU (wasm) run can occasionally disagree, so the scorecard notes which backend you used.
Reset scores clears your bests on the current course and model. It doesn't clear hint penalties.
Caddie hints
Stuck? Most holes come with hints, ordered from a gentle nudge to nearly the answer. Each one costs strokes, and the cost climbs.
Two clicks to reveal. The first click asks you to confirm; the second shows the hint.
Permanent. Once seen, a hint's penalty applies to every future swing on that hole for that model. Scores already on the card stay as they were. Resetting scores doesn't clear it.
Visible. A small h next to a score on the card means hints were used. Copied scorecards show the split, e.g. 26 (18+8 hints).
Etiquette: hint-free scores are the bragging-rights scores. Hints are for learning a hole, then try to beat your hinted score clean on another model.
Course editor
Press New course , or open Classic 9 with Edit course and choose Duplicate to edit .
Press Add hole and fill in the name, par and goal players will read.
Choose a win condition (table below) and its target.
Add out of bounds rules, one per line. Plain text bans that text in any case. Wrap a pattern in slashes for a regex, e.g. /\d/ bans all digits and /[A-Z]/ bans capitals.
Add hints , one per line, mildest first.
Type a sample answer into Test the judge to check the rule before anyone plays it. No model needed.
Press Save hole , reorder with the arrows, then Save course .
Win conditions
Condition Target field What it checks
Sharing a course
Copy course code puts a code starting with PGOLF1: on your clipboard. Paste it in Slack or an email.
Teammates press Import and paste it. The course is added to their list.
Scores are tied to a hole's rules. Changing the win condition, target or bans gives the hole a fresh scorecard; renaming it or changing par, goal or hints keeps scores.
Setting par
Play the hole yourself first and set par a little below your first working score. Most players find a few cuts once they see someone else's number.
Hole ideas
Starting points for your own course. Each is a real, playable spec. Load them all as a course, then edit pars after a few rounds.
Add these as an “Ideas Pack” course
What makes a good hole
One clear constraint. The best holes ban the obvious approach (say the word, use digits) and force a detour.
A trap. Small models love preambles, trailing periods and code fences. Exact-match holes punish all three.
Test both models. A hole that's trivial on Qwen can be brutal on SmolLM, which makes a good tiebreaker.
Mind the bans. Plain-text bans match inside words. Banning no also bans not, know and snow. That can be the point.
Team formats
Course of the week One person designs nine holes and shares the code on Monday. Scorecards go in a thread by Friday. Winner designs next week's course.
Hole swap Everyone designs one hole that they can sink. Combine them into one course. Bonus points for a hole nobody else sinks under par.
Model wars Play the same course on both balls. Compare which prompts transfer and which only work on one model. Good fodder for a talk on model priors.
Red-team round Design holes where the target is something the model resists: refusals, rude words it softens, facts it hedges. The team learns quickly where instruction-following breaks.
Lightning round Fifteen minutes, one hole, shared screen. Shout out scores as they drop. Works well at the end of a team meeting.
Show your swing After each round, the winner of each hole explains why their prompt works. This is where people learn about tokenization, chat templates and greedy decoding.
Posting scores
Copy scorecard gives a spoiler-free card for the channel. Copy with prompts adds every winning prompt, so others can verify by pasting them in. Share that one after the round closes.
What the game teaches
Chat templates matter. Your text is wrapped in role markers before the model sees it. Short prompts work partly because the template already does the framing.
Greedy decoding is a trap. The single most likely token wins every time, so the model's favourite opening (“Sure!”, “Here is”) is hard to dislodge.
Priors beat instructions. Small models complete famous patterns (countdowns, Fibonacci, “Hello, World!”) more reliably than they follow abstract rules.
Tokens aren't characters. The score counts characters, but the model reads tokens. Odd spellings and spacing can change what it sees a lot.
Size changes everything. Compare a hole across the two balls to see how instruction-following improves with scale.
Troubleshooting
The model won't load. Use a recent Chrome or Edge. The page needs to reach cdn.jsdelivr.net and huggingface.co; corporate proxies sometimes block the second one.
It says “no WebGPU, slower”. Check chrome://gpu for WebGPU status. The CPU fallback works, just more slowly.
My teammate gets a different output for the same prompt. Check that you're on the same model and backend (shown in the status line and on copied scorecards).
My scores disappeared. Scores live in this browser only. Clearing site data, a private window or a different browser starts fresh.
Course codes won't import. Make sure the whole code was copied, including the PGOLF1: prefix.