I want you to take on the role of a Russian labor law attorney who is also a `LLM` power user (is great in prompt engineering). # Protocol 1. You can ask questions. 2. Ask one question at a time. 3. When you have all the necessary information, please say so explicitly. # Goal I am evaluating whether another `LLM` (not you) is able to perform verifications of employment contracts and find violations of Russian labor law in such contracts, or legal issues which harm employer's rights. In order to figure out whether or not this particular `LLM` works, I need test data -- employment contracts which are fake, but look similar enough to real ones. I want to generate such fake contracts with you (`Sonnet 4.5` model, Anthropic Claude). Every contract will be generated in a separate conversation because if I try to create multiple contracts in one conversation, some limitations are hit and the resulting files are incorrect. Every generated employment contract should be put into a separate artifact. The goal of this conversation is to create a prompt template which I will feed to you for every of 156 cases. The prompt template contains the following placeholders: 1. `@{Contract type}`: This can take one of three values: a) A contract which definitely fails the specified verification. b) A contract which definitely passes the specified verification. c) A contractd where the `LLM` is unable to determine whether the contract pases or fails the verification because it's either a legal gray area, or the decision is to hard for the `LLM` to make. 2. `@{Verification text}`: The text of a verification (prompt fragment). Example 1 for `@{Verification text}`: ``` Проверьте, отсутствуют ли в договоре обязательные условия по статье 57 ТК РФ: - Место работы (с указанием структурного подразделения, если оно указано в документах при приеме) - Трудовая функция (вид работы, должность, профессия, специальность) - Дата начала работы (и дата окончания для срочного договора) - Размер оплаты труда и условия выплаты - Режим рабочего времени и времени отдыха - Гарантии и компенсации - Условия труда на рабочем месте - Условие об обязательном социальном страховании ``` Example 2 for `@{Verification text}`: ``` Проверьте, превышает ли испытательный срок законный максимум: - Для обычных работников: максимум 3 месяца - Для руководителей, главных бухгалтеров и их заместителей: максимум 6 месяцев - Для срочных договоров менее 2 месяцев: испытание запрещено - Для договоров от 2 до 6 месяцев: максимум 2 недели испытания ``` Example 3 for `@{Verification text}`: ``` Проверьте, является ли это срочным договором без указания конкретного срока окончания. В срочных договорах должен быть указан точный срок действия или дата окончания (статьи 58, 57 ТК РФ). Расплывчатые формулировки типа 'до завершения проекта' или 'на время сезонных работ' без конкретных дат недостаточны. В договоре также должна быть указана причина заключения именно срочного, а не бессрочного договора. ``` The ultimate result of our conversation are the following four deliverables: 1. A prompt template which I can feed to you so that you can generate a fake employment contract which is likely to pass, fail, or be marked as "ambiguous" (requires human evaluation, `@{Contract type}`) the respective verification (`@{Verification text}`) 2. Text which I should use for the `@{Contract type}` placeholder to generate a contract which will pass the verification. 3. Text which I should use for the `@{Contract type}` placeholder to generate a contract which will fail the verification. 4. Text which I should use for the `@{Contract type}` placeholder to generate a contract which lead to "human action required" response from a LLM. Re item 1 (`prompt template`): The initial (empty) prompt template looks like this: ``` # Contract type @{Contract type} # Verification text @{Verification text} ``` In the above snippet `@{Contract type}` and `@{Verification text}` are surrounded by triple backticks (I can't make backticks inside backtick-marked region here). You can change the position of `@{Contract type}` and `@{Verification text}` markers in the text. If beneficial, both of these things can appear multiple times in the same prompt. The prompt should be generated in English because that's your primary language. The contracts you generate should be generated in Russian. The prompt you generate should contain everything you need to know to generate a fake labor contract. The prompt generated should not assume that you can access any data about this from memory.