Promptshelf Regex vs NL Race · 15 tasks · pick the winner
Regex vs NL Race · 15 tasks

Regex or LLM? Fifteen real text-processing jobs. Pick the winner.

For each task, pick whether a classic regex or a natural-language instruction to a modern LLM is the better tool. You get the answer and an honest rationale right after. Eight of the fifteen are regex-wins. Seven are LLM-wins. Timer runs from your first pick.

Task 1 / 15
Score 0
Streak 0
Timer 0.0s
View-only replay. You're looking at a shared run (…). Scoring is locked. Click Start over to run your own.

Nice run.

Fifteen tasks complete.

Score
0 / 15
Time
0.0s
Longest streak
0
Guide Claude Code slash commands How slash commands delegate to regex, tools, or LLM reasoning inside Claude Code. Kit Cursor Rules Starter Kit Regex-first and NL-first patterns bundled as ready-to-drop .cursor/rules files. Play Prompt Ladder Daily 5-rung sort: re-order prompt fragments by which ships better code first.

FAQ

How is "correct" decided?
Each task has one canonical winner chosen by the engineering trade-off: cost, latency, determinism, and robustness on the described input. If a task's answer were genuinely ambiguous, we wouldn't include it. The rationale is visible after you pick, so you can disagree with honest friction.
Is this anti-LLM?
No. Seven of the fifteen tasks are honestly won by natural-language prompting. The point isn't that regex is better; the point is that each tool has a lane. Picking the wrong tool costs money, latency, or reliability — often invisibly.
Is this anti-regex?
No. Eight of the fifteen tasks are honestly won by regex. Regex is the cheapest, fastest, most deterministic text tool in existence for bounded formats. LLMs are flashier; regex is reliable.
What does the share URL encode?
Two things: ?b=<15-char bitstring> where bit i = 1 if you got task i correct, and ?t=<seconds> for elapsed time. Nothing else is sent. We do not store your run anywhere.
Why only 15 tasks?
Fifteen fits on a typical screen without scrolling fatigue, gives a meaningful streak count, and packs cleanly into a 15-bit URL. Twenty would dilute; ten would under-sample.
How do I suggest a new task?
Email hansng1988+regex@gmail.com with the task text, the honest winner, and a one-line rationale. Gmail plus-aliases route natively — no new account needed on our end.
Does this page send any data?
No. Picks, score, timer, and share URL are computed entirely client-side. No analytics, no telemetry, no third-party scripts beyond Google Fonts CSS. Your picks stay in your browser unless you copy the URL and share it.
Can I replay?
Yes. Click "Start over" from the end screen, or remove the ?b and ?t query params. A replay resets the timer and clears all picks. Because tasks are in fixed order, re-runs can be compared directly against prior URLs.

This page useful? Tip the maker in BTC — no account, no signup, just paste.

BTC bc1qs04leape97ner4wqa98n94l9n0gv9aa84eg4ux