M Blog is introducing a new AI agent skill called soul-audit. This tool offers a structured method for agents to re-run or deepen their identity interview after the initial setup phase. It's designed for developers and users who need to refine an agent's persona without the direct overhead of hand-editing configuration files. The core mechanism of soul-audit involves recording your answers into a dedicated shared answer bank, specifically located at state/interview.json. From this central source, the tool then automatically re-renders all associated identity files, ensuring that any adjustments you make are consistently applied across the agent's entire identity profile. This approach maintains a coherent and up-to-date agent representation based solely on confirmed user input.
How soul-audit Works
The process within soul-audit is compartmentalized into six distinct, independent phases. These phases cover critical aspects of an agent's persona: identity, communication style, mission, profile, access boundaries, and operational cadence. Users engage by providing direct answers to a series of questions presented within each of these specific categories. This modular design allows for focused adjustments without requiring a full re-evaluation of every aspect of the agent's identity each time a change is needed.
User inputs and answers do not directly write to files; instead, they flow through a controlled bootstrap interview command. This command supports three primary actions: set for inputting new answers, show for reviewing current answers, and confirm for finalizing the recorded input. Once answers are confirmed by the human user, the system then executes a render command. This command is responsible for processing the information stored in the shared answer bank and updating various identity files, notably SOUL.md and USER.md. This structured flow ensures that changes are deliberate and verified before being applied to the agent's operational identity.
Ensuring Data Integrity and Consistency
A fundamental design principle of soul-audit is its commitment to data integrity and consistency. All user-provided answers are recorded verbatim into a single, shared answer bank located at state/interview.json. This approach establishes state/interview.json as the definitive single source of truth for the agent's entire identity. Any information required for the agent's persona, from its mission statement to its communication preferences, originates from this central repository.
It's important to understand that the system does not generate any identity prose or invent content. The design actively prevents the system from fabricating or embellishing details. Instead, the tool maintains a strict adherence to user input. The system is engineered to structurally refuse to render or update identity files until all required answers have been explicitly provided by the human user, read back for verification, and then confirmed. This stringent confirmation step ensures that every piece of information stored in the shared answer bank remains precisely as the user input it, without any algorithmic interpretation or modification. Consequently, files such as SOUL.md and USER.md are guaranteed to re-render consistently, always reflecting only the human-confirmed verbatim input from the state/interview.json.
Practical Agent Identity Refinement
The utility of soul-audit becomes evident when an agent's initial setup requires modification or deeper definition. For example, consider an agent initially configured for a specific task. Over time, its operational parameters might evolve, or there could be a need to clarify its interaction style. Rather than manually navigating and editing multiple markdown or configuration files, a user can initiate the process through a simple prompt. The tool is designed to respond to a range of natural language triggers.
Typical triggers that activate the soul-audit flow include phrases like 'soul audit', 'customize agent', 'who am I', 'set up identity', or 'change my agent's personality'. When one of these triggers is detected, the agent initiates the identity interview. For instance, if you decide to 'change my agent's personality', this skill will guide you specifically through the communication style and identity phases of the interview. You provide new answers, which are then stored in the state/interview.json file. Once confirmed, the system automatically re-renders all relevant identity files, ensuring your agent's "personality" is updated without direct file manipulation.
The tool itself is implemented as a shell (mutating) utility. This classification is important because it signifies its capability to modify the underlying state of the agent. Specifically, it updates the state/interview.json file and subsequently triggers the re-rendering of dependent identity documents. This shell-based operation provides a robust and direct method for identity management within the agent's environment.
Frequently Asked Questions
Q: What if I don't answer all the required questions during the interview? A: The system is designed to structurally refuse to render identity files until all necessary answers have been provided and confirmed by the human user. You will be prompted to complete the interview.
Q: Does this skill generate creative prose or invent content for my agent's identity? A: No, this skill does not generate identity prose or invent content. Every answer recorded and used in the identity files is verbatim user input. The tool ensures your exact words are used.
Q: How does soul-audit ensure consistency across various identity files?
A: The single shared answer bank, state/interview.json, acts as the sole source of truth for all identity data. This means that all related files, such as SOUL.md and USER.md, consistently re-render from this verified central repository.
soul-audit provides a practical and consistent method for managing and evolving your AI agent's identity. It streamlines the update process by centralizing user input and automating file rendering, ensuring agent personas remain accurate and coherent.
