Define Clear Criteria and Evaluation Factors

For AI Functions, it’s important to provide specific judgment criteria and clearly indicate what factors should be considered in the evaluation.

Good Example (Sentiment Score):

For each provided sentence, calculate a polarity score in the range of -1.0 to +1.0.

Scoring Criteria:

Extremely Positive (+0.8 ~ +1.0)
- Expressions showing strong satisfaction or praise
- Situations where problems are completely resolved
- Mentions of outstanding performance or quality

Positive (+0.4 ~ +0.7)
- Mentions of clear advantages or strengths
- Reports of good experiences or results
- Results exceeding expectations

...(omitted)...

Extremely Negative (-1.0 ~ -0.8)
- Serious problems or malfunctions
- Strong dissatisfaction or negative emotions
- Major obstacles or hindrances

Factors to Consider in Assessment:
- Strength of expressions (degree adverbs like "very," "extremely," etc.)
- Degree of problem resolution (complete resolution, partial resolution, etc.)
- Relationship between expectations and results (above expectations, as expected, below expectations)
- Specificity and importance of advantages/disadvantages
- Overall context and intent

Rather than vague instructions, clearly listing the score ranges, judgment conditions, and factors to consider enables AI to analyze from multiple perspectives, significantly improving judgment accuracy.

Clearly Specify Output Format

By specifying what format you want AI to return results in, you can get results in your desired format.

For returning numeric values only:

Predict the virality of the provided social media post with a score from 0-100.

Evaluation criteria:
- Elements that evoke emotions
- Topicality/timeliness
- Elements that make people want to share
- Call-to-action prompts

Return only the score as a numeric value.

For comma-separated returns:

List 5 well-known American companies for each American industry classification.
Return multiple company names separated by commas for each industry.
Return the official company name.

Example: A, B, C, D, E

For specific format returns:

Extract the customer's specific request in one sentence from the provided email body,
and score the business impact from 1-10.

Output format: Request content | Score
Example: Add batch invoice generation feature | 8

Provide Concrete Examples

By including concrete examples in your prompts, AI’s understanding deepens and success rates for similar patterns increase.

Good Example:

This is name and email information.
Based on the name and email information, please estimate and return the correct "last name" 
in "Japanese" and "kanji." Even if the original information is in romaji, the result you 
return must be the last name in "kanji."

Example: Kato Kota -> 加藤

If you don't know, please return "Unknown."

Concrete examples are especially effective for:

• Format conversions (romaji → kanji, unit normalization, etc.)
• How to assign classification labels
• Name order arrangement

Specify Handling of Edge Cases

Specify in advance how to handle unknown cases and special cases.

Good Example:

You are a specialized assistant for identifying regions from school names.
Based on the provided school name, please determine whether the school is located in the 
Kanto region (Tokyo, Kanagawa, Chiba, Saitama, Ibaraki, Tochigi, Gunma).

If the school is in the Kanto region, answer "Kanto," if outside the Kanto region, 
answer "Non-Kanto."
If the school name is unclear, answer "NA."

This provides an “escape route” for AI when it’s uncertain, preventing errors and inappropriate guesses.

Make Classification Labels Specific and Comprehensive

When performing label classification, list all label candidates in advance.

Good Example:

Please assign labels to categorize the provided text into the following groups:

Good Support
Bad Support
High Product Rating
Low Product Rating
High Rating for Features
Low Rating for Features
Comparison with Competitor Services
Implementation-Related Content
Pricing-Related Content
Other

Always include an “Other” category to handle cases that don’t fit into any of the defined categories.

Specify Official Names and Notation Rules

When handling company names or proper nouns, clearly define notation rules.

Good Example:

Return the official legal entity name for each organization name.

From the provided investor data, correct notation inconsistencies and standardize 
company names as much as possible.

Example: 
Softbank -> Softbank
Softbank Group -> Softbank
Softbank Corp -> Softbank
Softbank Capital -> Softbank

Sequoia Capital -> Sequoia Capital
Sequoia Capital China -> Sequoia Capital

Cases Where AI Prompts Are Suitable

AI Prompts are suitable for the following types of rule-based processing and calculations:

Concrete Examples:

• “Determine whether sales are 1000 or more”
• “Create age group categories (teens, 20s, …) from age”
• “Label as ‘Order Required’ if inventory is 10 or less, otherwise ‘In Stock’”

For example, suppose you want to create an age group column from an age column.

Even with AI Functions, it’s possible to create an age group column for such cases.

However, using AI Prompts allows you to create the age group column faster and with better reproducibility.

Since AI Prompts return R scripts, they excel at returning labels based on conditions and calculations. On the other hand, when you want to return labels from text or product names—cases requiring understanding of context based on individual values—AI Functions are more suitable.

When using AI Prompts for labeling and calculations based on conditions, there are advantages in the following areas:

Performance

Because they execute as R scripts, there are no row limitations compared to AI Functions. Furthermore, since they don’t return values one by one, they are superior to AI Functions in terms of execution time and cost.

Reproducibility

Because they use conditional expressions in R scripts, the same data will always return the same results. With AI Functions, since AI judges the labels, slight variations may occur depending on the data or the task being instructed.

For more details on AI Prompts, please see this note.