Generating 156 Interpretation Texts with AI — A Strategy for Large-Scale Content

 

78 cards, each with upright and reversed orientations. That is 156 interpretation texts needed. Each one must be a natural 3-4 sentence paragraph that maintains the distinctive tone of divination. Writing all of this solo? Not realistic. But can you just tell AI "make me 156 interpretations" and call it done? Also no.

This is the story of why "framework first, generation second" matters so much in large-scale AI content creation, and how I validated the results.

Framework First, Generation Second

The most common mistake when requesting large volumes of text from AI is simply saying "just make it." If you ask "generate interpretations for all 78 tarot cards," AI will oblige. But the output will lack consistency. The tone drifts between early and late cards. Cards with similar meanings end up with overlapping phrasing. Some interpretations run long while others are oddly brief.

The solution is to define a framework before generation. Specify what to write, in what tone, in what structure, at what length. Then request generation within that framework. Think of it as writing the recipe before you cook.

The interpretation text framework for Tarot Master had four axes: tone guidelines, length standards, per-suit thematic framework, and reversed interpretation philosophy.

Finding the Right Tone for Korean Interpretations

The first axis was tone. Tarot interpretations are not standard information delivery. They need the distinctive atmosphere of divination.

English-language tarot interpretations typically use second-person direct address: "You are entering a period of transformation." Translated directly into Korean, this becomes awkward. It reads like a translation rather than something natural, and it misses the feel of how fortune-telling actually sounds in Korean culture.

I thought about how fortune tellers actually speak in South Korea. They do not start with "you." Instead, they use advisory constructions: "This is a time of new beginnings," "Caution is advised," "Now is the moment to seize the opportunity." The subject is implied, the situation is described, and advice follows.

I explicitly instructed the AI on this tone. Use advisory sentence endings. Avoid second-person pronouns. Ensure sentences flow naturally from one to the next.

A single prompt guideline shaped the tone of all 156 texts. Without this, the AI would have generated something that read like translated English tarot interpretations — technically correct but culturally off. This is where the framework proves its value.

Reversed Interpretation Philosophy: Not the Opposite, but the Shadow

The second critical definition was the philosophy of reversed interpretations. This had to be decided before any data generation began.

In tarot, there are broadly three schools of thought on how to read reversed cards. One treats them as the opposite of the upright meaning. Another sees them as the upright energy weakened or internalized. The third views them as the negative aspect of the upright meaning amplified.

The first approach (simple opposite) is intuitive but lacks depth. If the reverse of "success" is "failure," then having two orientations adds no real value. You end up with something as predictable as flipping a coin.

The second approach (weakened/internalized) is nuanced but too vague. Saying "The Magician's power is diminished" does not give the user much to work with.

Tarot Master primarily adopted the third approach. The guiding question: "If this card's energy were excessive or distorted, what problems would emerge?"

A concrete example: The Magician upright means "creativity, willpower, focus." Reversed does not mean "incompetence." It means "deception, wasted talent, manipulation." The ability exists, but it is being directed the wrong way. The Empress upright means "abundance, nurturing, nature." Reversed is not "poverty" — it is "overprotection, dependence, creative stagnation." The energy of abundance has warped into excessive caretaking or inertia.

I explained this philosophy to Claude as "the reversed meaning is the shadow aspect of the upright." The AI picked up on this context and generated text that reflected the nuance well. It was a moment that confirmed even abstract concepts can be effectively communicated to AI.

Major Arcana: AI Generation + Full Validation

With the framework defined, generation began. Major Arcana first.

The 22 Major Arcana are the heart of tarot. From The Fool to The World, each card carries powerful symbolism. They are also the cards users encounter most frequently. So I paid special attention to generation quality for these 22.

I requested all 22 interpretations in a single batch from Claude, providing the full framework: tone guidelines (advisory style, no second person), reversed interpretation philosophy (energy distortion), keyword criteria (around 4, concrete states), and interpretation length (3-4 sentences).

The quality was honestly impressive. The Fool's upright interpretation read: "A new journey begins. It is time to step into the unknown without fear. Embrace new possibilities with an open heart." The tone was right, the length appropriate, and the atmosphere genuinely felt like tarot.

But I did not use the output without full validation. I read through all 22 cards one by one. Three checkpoints: Does this contradict the traditional meaning of the card? Is the language natural? Do any interpretations overlap between cards?

Most corrections were in the "naturalness" category. AI-generated text is grammatically correct but occasionally has a subtle artificiality — a "nobody actually says it this way" quality. For instance, changing a formal passive construction to a simpler active one. Same meaning, but the latter reads more naturally.

Minor Arcana: Per-Suit Thematic Framework

The real challenge was the Minor Arcana. Four suits, 14 cards each, upright and reversed combined totaling 112 interpretation texts. Full validation like the Major Arcana was physically burdensome at this volume.

Here, I changed strategy. Before generation, I defined a per-suit thematic framework.

Each suit has its own thematic territory. Wands cover passion, creativity, and action. Cups cover emotion, relationships, and intuition. Swords cover intellect, conflict, and truth. Pentacles cover material matters, practicality, and achievement.

Within each suit, there is a narrative arc by number. Energy unfolds from Ace to King. Ace is the initial spark, Two is planning and decision, Three is early expansion, Four is stability and stagnation, Five is conflict and challenge, Six is recovery and balance, Seven is perseverance and strategy, Eight is transition and momentum, Nine is near-completion, and Ten is the final conclusion. Court cards (Page, Knight, Queen, King) represent the learner, the doer, the steward, and the master, respectively.

I explained this framework to Claude first, then requested generation in batches of 14 per suit. "The theme of the Wands suit is passion and action, and the narrative flows from Ace to King like this. Within this context, generate upright and reversed interpretations for all 14 cards."

The framework's effect was immediate. Consistency across cards held up. The problem of similar phrasing repeating within a suit diminished. Wands' Five addressed "competition and conflict" while staying within the Wands' broader theme of "passionate action." Swords' Five dealt with the same "conflict" motif but in the context of "intellectual dispute and argument" — Swords' territory.

Give AI sufficient context, and the consistency of the output improves. This was the principle I felt most strongly throughout the project.

Structural Sampling for Validation

Full validation of 112 Minor Arcana interpretations was not practical. But skipping validation entirely was unsettling. The solution was structural sampling.

Rather than picking random cards to check, I selected structurally important points for validation.

I validated the Ace of each suit. Aces represent the starting energy and core essence of their suit. If an Ace's interpretation does not align with the suit's theme, the entire flow is off. I validated the Five of each suit. Fives traditionally represent "conflict." I checked that all four suits address conflict but in different domains. I validated the Ten of each suit. Tens are completion. Checking where each suit's narrative concludes reveals whether the overall arc is coherent.

Court cards (Page, Knight, Queen, King) were validated separately. Court cards are interpreted differently from number cards. Number cards describe "situations" while court cards describe "personality types." I needed to confirm this distinction was properly reflected.

This sampling strategy covered 7 cards per suit, 28 cards total — half of the 56. More efficient than full validation while ensuring structurally critical points were not missed.

Issues Found During Validation

The sampling validation did surface problems that needed fixing.

The most common issue was "repetitive phrasing." AI tended to repeat similar sentence structures within a suit. Three consecutive cards might start with the same pattern, or several cards might end with the same closing phrase. These were manually reworded.

The second issue was "uneven depth in reversed interpretations." Some cards had reversed interpretations that were notably shorter than the upright or that read as simple opposites. These were either regenerated with the "energy distortion" philosophy re-emphasized or manually edited.

The third issue was "subtle unnaturalness in language" — the same problem found in the Major Arcana. Occasional stiff phrasing that is grammatically correct but sounds like no human would actually write it that way.

Despite these issues, the overall quality of the AI-generated first draft was at about "80% level." The difference in effort between starting from zero and starting from 80 and polishing to 100 is enormous. AI was the tool that delivered that 80%.

The Power of Prompts for Maintaining Consistency

The biggest takeaway from this process was the importance of prompts. The same AI given the same request produces dramatically different results depending on whether a framework is included in the prompt.

Without framework: "Generate interpretations for the 14 Wands cards." Result: interpretations exist, but the narrative connection between cards is weak and the tone is inconsistent.

With framework: "The Wands suit theme is passion and action. The energy progression from Ace to King follows this arc. Use advisory tone ('this is a time of...'). Reversed interpretations explore energy distortion. Four keywords each. Three to four sentences per interpretation." Result: consistent tone, narrative connection between cards, appropriate length.

This difference looks like "gave specific instructions vs. didn't." But the essence is deeper. Defining a framework means clarifying "the standards I want the output to meet" before generation begins. When the standard is clear, AI delivers accordingly. When the standard is vague, AI delivers vaguely.

This is a universal principle of AI collaboration. Whether generating code, writing documentation, or producing designs — to improve the quality of AI output, first improve the quality of the request. The most effective way to improve request quality is to define the framework first.

156 Texts, Complete

Through this process, all 156 interpretation texts were finished. The 44 Major Arcana texts (22 cards x upright/reversed) went through full validation. The 112 Minor Arcana texts went through structural sampling.

Throughout the process, a natural division of labor emerged: AI generates the first draft, humans define the framework and validate quality. AI takes the work from 0 to 80; humans take it from 80 to 100. Without this partnership, producing 156 interpretation texts would have been unrealistic.

What I Learned

First, the key to large-scale AI content generation is "framework first, generation second." Defining tone, length, theme, and philosophy before requesting generation dramatically improves consistency.

Second, when full validation is impossible, build a structural sampling strategy. Selecting structurally important points (Ace, Five, Ten, court cards) for validation is more efficient than random sampling.

Third, the most common issues in AI-generated content are "repetitive phrasing" and "subtle unnaturalness." The content is not wrong — AI just struggles to say the same thing differently and occasionally uses expressions no human would choose.

Fourth, prompt quality determines output quality. This is the most universal principle of AI collaboration. "Do your best" gets vague results. "In this format, at this length, in this tone" gets what you actually want.

Next Up

156 interpretation texts are ready. 78 cards' worth of data is complete. Now it is time to show these cards on screen. A tarot app's UI is different from a typical app. Atmosphere matters more than efficiency; immersion matters more than intuitiveness. In Part 6, I will cover how I translated the abstract requirement of "a night-sky sense of depth" into concrete hex codes and CSS.