A persona compiler, not a digital person
dot-skill takes a body of evidence about someone and produces instructions that an agent host can load as a reusable skill. It began as colleague.skill, aimed at retaining the working habits of a departing teammate, and now covers three families: colleagues, relationships, and celebrities or fictional characters. The repository is bilingual Chinese-English, with a full English README, install guide, and translated entry points for several other languages.
The useful idea is to separate more than tone. A colleague skill can contain both a Persona layer and a Work layer that records scope, procedures, output preferences, and decision habits. Relationship templates look for expression patterns, emotional triggers, conflict, and repair. Celebrity research uses works, interviews, decisions, external evaluations, and a timeline. Generated skills can be corrected through conversation, updated with new files, archived by version, and rolled back. That is more disciplined than asking a model to imitate someone from five adjectives.
It is still an imitation built from selected material. The repository's demos are persuasive illustrations, not evidence that the resulting skill predicts a real person's decisions. A generated profile may amplify a narrow period, mistake group-chat performance for private belief, or turn an author's repeated phrase into an absolute rule. The README itself calls the project a demo. Treat the result as an editable interface to recorded patterns, never as the person's testimony or authority.
Source collection is where the real cost lives
The manual path is approachable. Clone the repository into the supported host's skill directory, invoke /dot-skill, choose colleague, relationship, or celebrity, and supply an alias, description, and source. Claude Code, Codex, Hermes, OpenClaw, and DeepSeek Harness have documented locations or installers. PDF files, screenshots, email archives, Markdown, pasted text, Feishu exports, and WeChat database exports can feed the process. Python helper scripts support research cleanup and generated-skill management.
A meaningful profile needs curation. The project correctly ranks long-form first-person writing and decision records above casual chat or second-hand commentary. Users must remove irrelevant conversations, distinguish what the subject wrote from what others wrote about them, and check whether the selected period represents the person fairly. More text is not automatically better evidence.
Automatic collection raises the setup burden further. Feishu requires an app bot in relevant group chats. Slack requires an administrator-installed bot, and its free plan limits history to 90 days. DingTalk lacks a message-history API, so the project describes a browser route; open issue #105 directly questions whether that route works. WeChat data has to be exported first through separate tools, and issue #127 remains an open request around decrypting backup data.
Privacy is the deciding weakness
This project asks for some of the most sensitive material a developer can handle: private messages, work documents, email, family conversations, and credentials for collaboration platforms. The main README explains source quality but does not present an equally visible consent, minimization, deletion, or retention workflow. A company should not interpret technical access to Slack or Feishu as permission to build a behavioral replica of an employee. A personal user should not assume that participation in a chat grants permission to model another participant.
Two open issues make the operational warning concrete. Issue #113 reports that auto-collector setup uses ordinary terminal input for tokens and passwords, allowing secrets to appear on screen and in session recordings. Issue #112 reports that collector configuration files can inherit permissive modes such as 0644 rather than enforcing owner-only access. Pull requests propose masked prompts and 0600 files, but both issues remained open in the activity inspected for this review. Until the fixes land and are verified, use manual sources or isolate the collectors on a private account with restricted tokens and a strict file umask.
Generated skills need protection too. They distill the input into compact statements that may be easier to browse, copy, or accidentally commit than the original archive. Keep them out of shared repositories, inspect host backup and synchronization behavior, and document who can invoke them. Deleting the source files does not remove conclusions already written into a generated skill.
The architecture is clever, with visible seams
The repository follows the AgentSkills directory format, so it can run as instructions plus local Python tools instead of requiring a hosted backend. Users can inspect the prompts, generated Markdown, version archives, collectors, and installers. The MIT license permits modification, and the technical report explains the earlier colleague-specific architecture.
The family expansion is newer than that paper, however, and the split outputs are not perfectly isolated. Issue #134 shows that a work-only skill can still instruct the model to respond in the colleague's style and refer to a Persona section absent from the artifact. This matters for teams that deliberately want documented expertise without impersonation. It also demonstrates a general limit of prompt-built products: labels such as work-only are behavioral intentions until tests verify the exact generated text.
Health and the recommendation
The last push was August 13, 2026, and an issue about WeChat export was updated August 14. The repository has 34 open issues and pull requests combined, more than 22,000 stars, and no formal GitHub release. Recent commits added host support and documentation, while the most recent non-documentation commit in the inspected history was an April compatibility fix. That is active maintenance, but not a conventional release cadence.
The documentation is energetic and unusually visual, with a source matrix, host paths, examples, project layout, technical report, and multilingual links. It is weaker where the consequences are highest: privacy and credential guidance should be as prominent as persona quality.
dot-skill is best approached as a transparent research and prototyping tool. Use your own writing, public sources, fictional characters, or material covered by explicit consent. For organizational knowledge transfer, write down the governance before collecting anything. If all you need is repeatable work instructions, Anthropic's skills examples avoid the imitation problem; if you need a long-lived agent with memory, Mem0 or Eliza addresses a different and often safer requirement.