Anthropic Fable 5.1 Usage Guide

When a coding request is rejected, the questioning method should be changed from "compile without errors" to "there's a bug in this code." Lee Jong-beom, an adjunct AI professor, emphasized that if the language is not well-known, it is advisable to attach the official documentation of that language for grammar reference.

He added that if the model is returning Base64 data, excluding Base64 from the results delivered by the model can prevent the problem of requests being rejected.

Starting with the default High setting, compare the same task with Medium or Low, and if the results are sufficient, using lower settings is more cost-effective.

The Low setting may not retrieve enough up-to-date information, so additional prompts should be used to encourage the retrieval of missing information. High or Max settings can sometimes consume the entire document; if a high setting must be used, it's effective to request step-by-step output by adding 'only write the final document as a final answer when I ask you to' at the end of the request.

If the AI misses small numbers in dense charts, actions must be taken to magnify the necessary parts for it to re-examine. Simply instructing it to "magnify" is insufficient; a tool that actually crops and magnifies images should be introduced.

Taking a sales graph as an example, it is effective to select the area for the AI to check and configure the tool to crop that part and return it as a magnified image. Image cropping processes can be implemented using libraries like OpenCV.

This prompt encourages the AI to remove embellished expressions and explain meanings directly.

While this prompt can be included in the system prompt, direct injection is primarily recommended.

This contributes to reducing the average completion time while maintaining similar quality and cost.

If parallel processing is deemed possible, the Steering function can be used to enforce it.

It is important to prevent the AI from modifying unrelated screens even when asked to perform a specific task, such as resolving a login error. Adding the following prompt is effective for this purpose:

'If you find issues unrelated to the request, do not fix them and inform me after the task is complete. Please proceed with any modifications strictly necessary to resolve the requested issue.' This prompt encourages the AI to focus only on the requested problem and report other matters.

This prompt can also be useful when writing tests, contributing to the accuracy and efficiency of work by preventing unnecessary AI intervention.