dot-skill — Digital Life Distiller
Distill anyone—colleagues, loved ones, idols—into an interactive AI Skill that thinks and speaks like them.
Evidence shows: tools require user-provided data sources (e.g., chat logs, documents) but do not clearly state how data is processed, stored, or transmitted; auto-collectors (Feishu/DingTalk/Slack) require credentials but no least-privilege or user-confirmation mechanisms are described; version rollback exists (version_manager.py) but completeness and data retention are unclear; dependencies are not pinned, posing supply-chain risk; external effects (e.g., installing to multiple hosts) lack explicit user confirmation. Deductions: lack of data-flow transparency and user confirmation, insufficient dependency security.
Evidence shows: clear project structure, README consistent with code; tests cover CLI lifecycle and installation but not all tools; dependencies not pinned, affecting availability; error messages partially present in tests but not systematic. Deductions: dependency availability not guaranteed, error handling incomplete.
Evidence shows: explicit support for multiple hosts (Claude Code, Hermes, etc.) and data sources, scenarios cover colleague, relationship, celebrity; capability boundaries stated (e.g., DingTalk lacks message history); trigger commands clear (/dot-skill etc.); environment fit with install guides. Deductions: some features (e.g., photo sharing) marked as coming soon, limiting actual capability.
Evidence shows: clear information architecture with README, INSTALL.md, PRD; detailed install notes; stable naming (dot-skill); examples and FAQ (community gallery); known limitations mentioned (demo version); MIT license explicit; version management tool exists but no CHANGELOG; maintenance responsibility unclear. Deductions: missing CHANGELOG, unclear maintenance responsibility.
Evidence shows: output is usable Skill files with clear structure; high marginal value (solves specific problem); cost-benefit not fully assessed, multiple optional dependencies may increase install burden. Deductions: cost-benefit not quantified.
Evidence shows: README claims (e.g., 20K stars) not verifiable from source; technical report cites arXiv but no verification; tests exist but not independently verified. Deductions: claims lack traceability, facts and inferences not clearly separated.
- Data flow is opaque: users provide sensitive chat logs but no clarity on processing, storage, or transmission, posing privacy risks.
- Dependencies are not pinned, posing supply-chain security risks.
- Auto-collectors require credentials but no least-privilege or user-confirmation mechanisms are described.
- README claims (e.g., 20K stars) are not verifiable from source; treat with caution.
What does this agent do, and when should you use it?
dot-skill (formerly colleague.skill) is a skill-generation engine built on the AgentSkills open standard. It collects source material from multiple data sources (Feishu, DingTalk, Slack, WeChat chat logs, PDFs, emails, Markdown, etc.) and uses prompt pipelines to generate AI Skills with specific personas and professional capabilities. It supports three character families: colleague, relationship, and celebrity, each with tailored data-collection and analysis workflows. Generated Skills can be installed into multiple hosts: Claude Code, Hermes, OpenClaw, Codex, and DeepSeek Harness. The repo includes a full Python toolchain: Feishu/DingTalk/Slack auto-collectors, a celebrity research pipeline (subtitle download, transcription, research merge, quality check), multi-host installers, and version management. The technical report 'COLLEAGUE.SKILL: Automated AI Skill Generation via Expert Knowledge Distillation' is available.
dot-skill provides a unified /dot-skill entrypoint that guides users to select a character family (colleague, relationship, or celebrity), enter an alias, basic profile, personality tags, and choose a data source. For the colleague family, it automatically invokes Feishu, DingTalk, or Slack APIs (or browser-based collection for DingTalk) to gather messages, docs, and spreadsheets, then generates a Skill with a work skill and a 6-layer persona structure (SKILL.md, persona.md, work.md). For relationship, it analyzes chat logs for expression DNA, emotional triggers, conflict and repair patterns. For celebrity, it uses the six-dimension research toolchain (download_subtitles.sh, srt_to_transcript.py, merge_research.py, quality_check.py) to distill mental models and decision heuristics from subtitles, interviews, and decision records. Generated Skills are invoked via /{character}-{slug} and support incremental merging, conversation correction, and version rollback.
- A manager wants to preserve the technical judgment and workflow of a departing senior engineer by distilling their docs and review comments into a Skill for the team.
- Someone in a long-distance relationship wants to recreate their partner's texting style from months of chat logs for emotional companionship.
- A fan wants to interact with a simulated version of their favorite author using their books, interviews, and blog posts.
- A product team wants to distill a legendary designer's decision frameworks from public talks and Q&As into a design-review Skill.
- A person wants to preserve the memory of a deceased relative by converting their social media posts and diaries into a conversational tribute.
What are this agent's strengths and limitations?
- Automated data collection from Feishu, DingTalk, and Slack with just a name input, covering messages, docs, and spreadsheets.
- Three distinct character families with dedicated prompt pipelines and generation templates, making it flexible for different scenarios.
- Cross-host support for Claude Code, Hermes, OpenClaw, Codex, and DeepSeek Harness, with generated Skills installable across them.
- Celebrity family ships a six-dimension research toolchain (subtitles → transcript → research merge → quality check) that reproduces mental models, not just tone.
- Includes version control, incremental merging, and conversation correction for iterative refinement of Skills.
- DingTalk API lacks message-history support, requiring browser-based collection; Slack free tier limits history to 90 days.
- Skill quality depends heavily on source-material quality; low-quality or second-hand sources degrade output.
- Project is still a demo and may contain bugs; Feishu auto-collection requires adding the bot to relevant group chats.
- Photo-sharing for relationship family is not yet available (coming soon), so capabilities are currently limited to text-based interactions.
- Requires familiarity with multi-host installation paths and Python tools, which may be a barrier for non-technical users.
How do you install or deploy this agent?
In a supported host (Claude Code, Hermes, OpenClaw, Codex, DeepSeek Harness), tell your agent: 'Install the dot-skill skill for me: https://github.com/titanwings/colleague-skill'. The agent will clone the repo and register the entrypoint. For manual install, clone the repo into the host's skills directory: Claude Code → ~/.claude/skills/dot-skill, OpenClaw → ~/.openclaw/workspace/skills/dot-skill, Codex → ~/.codex/skills/dot-skill, DeepSeek Harness → ~/.dsh/skills/dot-skill (global) or .dsh/skills/dot-skill (project). For Hermes, after cloning run 'python3 tools/install_hermes_skill.py --force'.
How do you use this agent?
After installation, type /dot-skill or say 'start dot-skill' in the host. Choose a character family (colleague, relationship, celebrity), then enter alias, basic profile, personality tags, and select a data source. All fields are optional; even just a description can generate a Skill. Once created, invoke with /{character}-{slug} for full Skill, /{character}-{slug}-work for work-only, or /{character}-{slug}-persona for persona-only. Use 'python3 tools/skill_writer.py --action list' to list generated Skills and 'python3 tools/version_manager.py --action rollback' to roll back.
How does this agent compare with similar options?
It shares a community case with alchaincyf/karpathy-skill (a Karpathy distillation), but no direct competitors are mentioned.