Preserve a colleague's judgment and tone as a reusable AI agent profile
When a colleague leaves or a mentor departs, it is often difficult to inherit their work methods and decision-making criteria. Distilly analyzes materials such as messages, documents, and interviews to create a reusable 'person profile' of that person's experience, decision patterns, and expression style. This profile is packaged as an agent skill, allowing it to be installed on various AI agent hosts for immediate use.
Expanding to three types: colleagues, partners, and public figures
The project started with creating work skills for colleagues but now supports three types of person profiles. The colleague type analyzes work standards, workflows, and expression styles to generate work skills and personas, with automatic data collection from Lark, DingTalk, and Slack. The relationship type organizes emotional triggers, conflict patterns, and reconciliation methods into a persona skill. The public figure type provides a toolchain for researching subtitles, interviews, and decision records across six dimensions. Each type has different data collection strategies, analysis dimensions, and profile structures.
Installation and data collection scope by agent host
Installation is simple; you can instruct the agent to install it or clone the repository and place it in the skill directory. Data collection automatically retrieves messages via API from Lark and Slack, while DingTalk supports browser-based collection. You can also manually upload various file formats, including PDFs, images, emails, and Markdown files. Grok Bot allows previewing stored skill workflows, but direct local installation verification has not yet been performed.
Profile generation structure and evolution method
The generated skill uses a persona as its base and adds type-specific modules. The colleague type includes a 6-layer persona and work skill modules, the relationship type covers expression DNA and emotional patterns, and the public figure type addresses mental models and decision heuristics. The profile is not static; it evolves. When new files are added, it analyzes differences and merges them into relevant sections without overwriting existing conclusions. If you correct the profile during a conversation by saying 'that person would not do that,' the change is immediately recorded in the correction layer and reflected. All updates are automatically versioned and archived, allowing you to roll back to previous versions.
Data quality determines profile quality
The tool's performance relies heavily on the quality of input data. For colleagues, their long-form writing or review comments are the best materials, while for relationships, long conversation records containing both conflict and reconciliation are ideal. For public figures, first-person books, blogs, and long interviews are prioritized, as secondary interpretations alone make it difficult to create an accurate profile. The project is currently classified as a demo version, and users are encouraged to register issues if they find bugs. It is released under the MIT license, allowing free use.