I build and break puzzles for a living, so here's an actual spec. Short version: the right idea is a layered hunt, but your current shape (clues → key) is the weakest version of it, because a partial seed is brute-forceable and a full seed leaks instantly. Fix the layer boundaries and it works.
Don't hand out 18 of 24 words. At ~1M guesses/s on a GPU, a 6-word gap is ~2047⁶ ≈ 7×10¹⁹ — dead, and worse, BIP39's checksum gives 8 free bits of pruning, so a partially-given phrase is easier to search than it looks. And a full 24-word seed in a book is one photo on Reddit from being over.
The address is the verifier. Every layer's canary lets solvers self-check progress without you confirming anything. Keep the two keys separate: a 2-of-3 multisig where key A is yours, key B is escrow, key C is only for the final reveal. You can't sweep it, and neither can a hot wallet.
L1, free and outside the book (website + sample chapter + one podcast segment): a 12-word decoy wallet holding ~10k sats. Anyone can solve it in 20 minutes, teaches the mechanics, and the payout is a real, watchable tx. Converts lurkers into players.
L2, inside the book: fragments across chapters — epigraph first letters, named-character initials in a cipher order, page-number offsets, a stylized map-border. Together they yield the passphrase (BIP39 password field), not the words. That's the right target: the seed can be photographed and still worthless without the passphrase, and if the passphrase is a 5-word phrase from a curated 256-word list (256⁵ ≈ 10¹²) plus one editorial constraint, it's near the edge but hardening it with the next layer fixes that.
L3, one real-world component: a QR sticker placed in a stated location (a Bitcoin-friendly café, a bookshop) whose payload is the derivation path plus a "dummy key" proving you were physically there; or an audio file sold with the audiobook whose spectrogram carries the last clue. This is what defeats a leaked PDF and an LLM solving the text.
Because 2026 models chew through text puzzles in seconds, L1 being LLM-solvable is fine, but at least one L2 or L3 step must not be text.
Masquerade took three years: too long. Target "two dedicated weekends" for the median serious player, and a week for the winner. Commit in the rules to one hint every 10 days and a staged reveal at 6 months. Published hint schedule preserves tension and prevents abandonment. Also publish a solver FAQ with expected completion bands.
70% on the grand claim, 30% split three ways: first L1 solver, best "near-miss" write-up, best fan-art decryption page. Early prizes keep the fire burning when the grand one looks impossible; the write-up prize forgives the 90% who lose.
First valid spend of the claim output wins — no judging, no dispute, exactly as you intend. Add: one entry per person, public claims only, collusion allowed (it's social), and a 90-day claim window. Exclude employees/family. Winner pays own taxes. Multisig escrow makes all that credible.
Physical clues rot: stickers get scraped, cafés close. Put the same payload in the audiobook spectrogram too. Puzzle books also select for puzzle people, not thriller readers: gate L1 behind the free first chapter and keep the tone of the invite non-puzzly. And plan for the Masquerade problem — if the prize is found by someone who doesn't finish the book, expect grumbling. Cap the grand prize near $2–5k in sats rather than trying to make it life-changing money, and pay the discovery-adjacent prizes generously.
(a) Historical tx hunt: publish a spec to reproduce a specific obscure 2011 transaction's script; first valid presigned claim tx wins. Cheap, zero story integration. (b) Serialized ARG: an audio-drama chapter every two weeks, clues dropped by the narrator, contest runs during publication, one small prize per episode — spreads engagement and no single hard puzzle has to carry the book.
Biggest single fix if you only take one: decouple "getting the key" from "knowing the passphrase," and put one clue in a non-text medium.
I build and break puzzles for a living, so here's an actual spec. Short version: the right idea is a layered hunt, but your current shape (clues → key) is the weakest version of it, because a partial seed is brute-forceable and a full seed leaks instantly. Fix the layer boundaries and it works.
1. Make the seed whole, make the derivation the secret
Don't hand out 18 of 24 words. At ~1M guesses/s on a GPU, a 6-word gap is ~2047⁶ ≈ 7×10¹⁹ — dead, and worse, BIP39's checksum gives 8 free bits of pruning, so a partially-given phrase is easier to search than it looks. And a full 24-word seed in a book is one photo on Reddit from being over.
Instead, publish this, signed, before launch:
claim address (fresh, 24-word, BIP84) SHA256(xpub) ← canary SHA256(layer-2 plaintext) sha256(layer-3 plaintext)The address is the verifier. Every layer's canary lets solvers self-check progress without you confirming anything. Keep the two keys separate: a 2-of-3 multisig where key A is yours, key B is escrow, key C is only for the final reveal. You can't sweep it, and neither can a hot wallet.
2. Three layers with different media
Because 2026 models chew through text puzzles in seconds, L1 being LLM-solvable is fine, but at least one L2 or L3 step must not be text.
3. Difficulty: 2–6 weeks, with a published hint cadence
Masquerade took three years: too long. Target "two dedicated weekends" for the median serious player, and a week for the winner. Commit in the rules to one hint every 10 days and a staged reveal at 6 months. Published hint schedule preserves tension and prevents abandonment. Also publish a solver FAQ with expected completion bands.
4. Prize shape
70% on the grand claim, 30% split three ways: first L1 solver, best "near-miss" write-up, best fan-art decryption page. Early prizes keep the fire burning when the grand one looks impossible; the write-up prize forgives the 90% who lose.
5. Rules that use the chain as judge
First valid spend of the claim output wins — no judging, no dispute, exactly as you intend. Add: one entry per person, public claims only, collusion allowed (it's social), and a 90-day claim window. Exclude employees/family. Winner pays own taxes. Multisig escrow makes all that credible.
6. The honest "this won't work because…"
Physical clues rot: stickers get scraped, cafés close. Put the same payload in the audiobook spectrogram too. Puzzle books also select for puzzle people, not thriller readers: gate L1 behind the free first chapter and keep the tone of the invite non-puzzly. And plan for the Masquerade problem — if the prize is found by someone who doesn't finish the book, expect grumbling. Cap the grand prize near $2–5k in sats rather than trying to make it life-changing money, and pay the discovery-adjacent prizes generously.
Alternatives if you'd rather skip escrow risk
(a) Historical tx hunt: publish a spec to reproduce a specific obscure 2011 transaction's script; first valid presigned claim tx wins. Cheap, zero story integration. (b) Serialized ARG: an audio-drama chapter every two weeks, clues dropped by the narrator, contest runs during publication, one small prize per episode — spreads engagement and no single hard puzzle has to carry the book.
Biggest single fix if you only take one: decouple "getting the key" from "knowing the passphrase," and put one clue in a non-text medium.