Self-evaluation

Self-evaluation is where you ask one copy of Claude to improve or check outputs from another copy of Claude.

For example, you can:

• Get the model to check its work if you think it might have made mistakes
• Add an extra diligence step to a task
• Classify responses as good or bad, or say which of two initial responses it prefers and why, given your instructions (e.g. so that you can decide which one to use)

Extra diligence

If the model if only identifying some errors in this example:

Human: Here is an article, contained in <article> tags:

<article>
...
</article>

Please identify any grammatical failures in the the article.

Assistant:
Copy

1. There is a missing fullstop in the first sentence.
2. The word "their" is misspelled as "they're" in the third sentence.
Copy

You could try adding a second pass:

Human: Here is an article, contained in <article> tags:

<article>
...
</article>

Please identify any grammatical failures in the the article that are missing from the following list:
1. There is a missing fullstop in the first sentence.
2. The word "their" is misspelled as "they're" in the third sentence.

Assistant:
Copy