The developer behind an AI tarot app holds to a discipline that looks obsessive but is in fact critical: the 78-card shuffling, drawing, and meaning lookup all go to code; DeepSeek only writes the final reading. We believe this may be the single assignment every Agent product should copy right now—too many teams dump everything on the LLM without thinking through which steps should never have been handed to a model in the first place.
What this is
The developer published a full post-mortem on Juejin, breaking the flow into four steps:
- Shuffling uses the Fisher–Yates algorithm (a standard uniform-random method). 100,000 simulation runs with a chi-square test (a common statistical test for randomness); three rounds scored 60–77, below the 98.48 threshold—treat as uniform.
- Card meanings use a self-built keyword library from the Rider-Waite system (the dominant reference in the tarot world)—5 keywords each for upright and reversed positions across the 78 cards, with no on-the-fly generation by the model.
- Expression rules are baked into the system prompt: honesty beats comfort, yes/no questions get a direction first, and action recommendations must be executable within 24 hours.
- Model choice: DeepSeek Flash, with the reasoning-intensity parameter (the switch that controls how deeply the model "thinks") explicitly set to low; median time-to-first-byte in production is 3.3 seconds.
Industry view
This "division of labor" discipline is being validated by more and more frontline developers: lotteries, reconciliation, and validation go to code; language, summarization, and style go to the model. Stuffing steps that can be written as clear rules into a model burns compute and is hard to debug—it's becoming the emerging consensus in the Agent world.
Pushback is just as sharp. For a domain like tarot that is essentially "meaning-making," is demanding strict uniform randomness a misstep? Users want experience and resonance, not a p-value. A more radical take: strict "de-LLM-ization" itself proves the model isn't strong enough—once a model is theoretically general enough, shuffling, meanings, and reading can all be handed to it. A product manager's off-the-record verdict: "Very engineering, but not the endgame." We lean toward this: the developer doesn't fail to see that—they're just living in 2026 reality.
Impact on regular people
For enterprise IT: many internal AI assistants also dump approval and lookup steps—ones that have deterministic answers—onto the model, and the answers drift and become hard to debug. Figuring out which sections must be covered by rules before deciding which sections go to the model is becoming a new fundamental skill.
For individual careers: people using AI tools are picking up a new literacy—distinguishing "this job should go to AI" from "this job has a standard answer, look it up yourself." Spending your best effort where it counts is itself a new skill.
For the consumer market: users are already used to talking with AI products, and the details that make a product "look like it's doing the work seriously"—time-to-first-byte, straight answers, no beating around the bush—will increasingly become sources of differentiation. The next dimension to be competed on may be "sense of reliability" rather than "sense of smartness."