Harfi gives you a hard spec, exactly three bullets, under 50 words, valid JSON with no code fence, and you write the instruction, under 120 words, that makes an AI model hit it exactly; you never write the answer yourself. Scoring is done entirely by a parser that counts bullets, counts words and checks JSON, so the same attempt always gets the same result, unlike an LLM judging how good your answer felt. Set 1, 8 trials, is free forever with no account; the other four sets, 40 trials total, are a one-time $29 purchase with a 14-day refund. Most trials run against Claude Haiku 4.5, deliberately not the strongest model available, so a perfect instruction-follower wouldn't make every trial trivial.
What holds up
- +Scored by a parser that counts and checks — the same attempt always gets the same result
- +Set 1 (8 trials) is free forever, no account or card needed
- +One-time $29 for all 5 sets and 40 trials — no subscription to renew
- +Deliberately uses Claude Haiku 4.5, not the strongest model, so the trials stay genuinely hard
Mind the limits
- −Only 8 of 40 trials are free — the other 32 need the $29 purchase to unlock
- −You write only the instruction, never the answer itself — not for practicing direct output writing