The obvious feature was not the durable one
The earliest version of Liminal Tarot I can reconstruct put AI interpretation close to the centre. That was the obvious moment of value: the user drew cards, and the system returned something useful.
It was also the most replaceable part.
That split has an ordinary version. A gym membership is what lets you book tomorrow, and any gym in the city can sell you one. The log of what you actually did over two years — which session, which weight, which week you stopped showing up — is not something a competing gym can hand you. When the membership lapses, you lose the booking. You do not lose the log.
Across the project history — from early March through June 2026 — the product became more coherent as AI moved into the infrastructure and continuity moved into the foreground. The lasting value was not a more impressive reading. It was the user's thread: the cards they drew, the notes they wrote, the questions they returned to, the Chapters that held a period of their life, and the expectation that the system would still remember all of it later.
Generated interpretation can be useful immediately. Continuity becomes useful because the person came back.
The product had breadth, not hierarchy
The early product already contained the seed of this idea. It supported physical decks as well as digital drawing, and readings could connect across time. Even then, Liminal Tarot was not only generating an answer. It was holding part of a practice.
But the first broad implementation expanded many surfaces at once: Daily Card, rotating prompts, physical- and digital-deck readings, Chapters, journaling, sharing, retention behaviour, and the supporting account systems around them.
The implementation was AI-assisted. The product decisions and accountability stemmed from years of "I wished this exists" as a tarot hobbyist.
That wishlist produced A LOT of features from the get-go, but it did not automatically produce a clear product thesis.
A compiling build could still be incomplete. A large feature set could still feel like a form wrapped around generated text. The product had breadth, but its load-bearing hierarchy was not yet clear.
It was never a perfectly linear strategy. There was plenty of research, implementation planning, and roadmapping before the build began, but the earliest pre-GitHub sequence is incomplete. I was only able to see that period more clearly in retrospect, using the Product Build History workflow to reconstruct the decisions that survived. What the record does support—and what the broader Liminal Tarot casefile now makes easier to trace—is the direction of travel: repeated use, retained context, and trustworthy state kept becoming more important than the novelty of any single interpretation.
Ritual was part of the mechanism
Ritual can sound like atmosphere added after the product works. In Liminal Tarot, it increasingly became part of how the product worked.
Onboarding shifted toward being the user's first real practice rather than a tutorial about the interface. Physical-deck support let someone keep the order and handling of their own cards while still saving the reading and receiving interpretation. Digital drawing needed movement, pacing, reveal timing, sound, and quiet loading—not because animation makes an answer more correct, but because the interaction needed a beginning, a moment of attention, and an ending.
An AI wrapper can return text after an input. A practice system has to support what happens before the text and what remains after it.
That same standard required restraint. Cached or offline states could not be presented as full capability when account and generation systems were unavailable. Stale state had to look stale. Actions that could not work had to remain unavailable.
The product had to preserve the practice without pretending the entire system was present.
Ownership had to survive entitlement
Chapters made the continuity thesis visible. They grouped readings around an active period rather than leaving each one isolated. Notes could live at card, reading, and Chapter level. Daily Card and saved readings supplied repeated inputs. Pattern language moved toward one consistent authority so different surfaces would not tell different stories about the same history.
The user was no longer only receiving an interpretation. They were building a private record.
That made entitlement decisions product decisions.
When a subscription changed, premium creation, editing, assignment, or synthesis could stop. Existing Chapters and their saved content did not disappear. The user could still list, view, read, detach, and delete what they had already made. The entitlement governed future capability. It did not revoke possession of the archive.
A separate change stopped free readings from disappearing after a limited retention period. Earlier retention logic had used expiry as conversion pressure. That is the gym deleting your training log to make re-joining feel necessary. The revised direction kept history visible and moved the commercial boundary toward creation limits and premium practice surfaces instead.
That correction mattered because the product promise and the data policy had to agree. A system cannot say it holds the user's thread and then erase the thread when the user does not upgrade.
Technical failures became continuity failures
Several failures looked technical until I applied the product thesis to them.
A stale native bundle could show an older experience even when the source changes were correct. Restore Purchases was not complete until Liminal Tarot's own state recognized the entitlement. Account deletion was not complete while the authentication identity remained. Daily Card could not reveal a final card that had not been persisted. A reading and its associated credit could not be allowed to split under retries. Pattern claims could not change because two screens counted history differently.
Each incident broke a different layer. The product failure was the same: what the user saw and what the system could preserve had diverged.
This is where the gym comparison stops. A paper log cannot contradict itself; a product assembled from several layers can, and every incident above was a version of that.
The response became increasingly contractual. Visible state had to correspond to authoritative state. Stored practice data had to survive entitlement changes. One rule had to govern recurring-card patterns. The reviewed candidate had to be the candidate that reached the installed product.
This is where continuity stopped being a content theme and became a quality standard.
The product could not claim to remember the user if its own layers disagreed about what was true.
Reflection was stronger than prediction
The language narrowed with the mechanism.
Earlier category assumptions still carried some predictive and mystical residue. Over the covered period, the product moved toward grounded reflection: cards as a mirror, private practice, daily return, physical or digital decks, history, Chapters, and notes. Predictive language and "Oracle" framing were removed.
Part of that was a review and store-positioning constraint. It also made the product more honest.
The strongest version of Liminal Tarot did not need to claim authority over the future, health, relationships, or major decisions. It could help someone examine what was present, save what they noticed, and return later with more context than they had before.
That fit the product better than a promise of certainty ever could.
AI became infrastructure
AI did not disappear from the product. Its role changed.
Interpretation and synthesis still mattered. They could make a reading legible, connect several readings, and support reflective prompts. But the durable value increasingly lived around the generated text: the ritual before it, the history after it, the Chapters that contextualized it, the notes that belonged to the user, and the ownership rules that protected the archive.
Looking back, the strategic question changed with that shift. The product was no longer only being shaped around whether the model could produce a better reading. The stronger requirement was whether the system could help someone continue a practice.
That is a less glamorous problem and a harder one. It requires retention without coercive erasure, entitlement boundaries without confiscation, platform-specific interaction without splitting the product, consistent data authority, and language that does not overclaim.
It also produces something a competing model call cannot recreate instantly. Another system can generate another interpretation. It does not inherit the private context a person has built over time.
What this reconstruction can support
This entry was reconstructed in August 2026 from accepted project history covering March 3 through June 10, 2026. It does not establish Liminal Tarot's current provider, release, deployment, revenue, user, traction, entitlement, or runtime state.
The earliest chronology remains incomplete. Some early corrections are supported through audit reporting rather than a complete independent reconstruction, and selected device and provider diagnostics and exact review timestamps are missing. The evidence supports the direction and causal structure of the product evolution. It does not support a claim that I understood the full strategy from the beginning, or that every step followed a neat sequence.
What I can support is narrower, and more useful.
AI could interpret the thread. The product had to keep it.
// End of transmission. Keep the thread. — ZYANE