How to Digitize a Departed Senior and Turn Them into an AI Agent Skill

Imagine a familiar scene. A lead backend engineer leaves for another company, taking with them an understanding of half the legacy services, the habit of rejecting every other pull request due to lack of context, and the tendency to write terse comments like "you've got an N+1 query here, redo it." Confluence documentation doesn't help because a living engineer makes decisions intuitively, relying on dozens of unwritten rules.

Developers from Shanghai AI Lab decided to tackle the problem literally. They wrote a framework dot-skill (formerly known as colleague-skill) that collects a person's texts, code, comments, and work correspondence, then packages them into a reusable AI skill following the AgentSkills standard.

Why Distill Living People

The project idea grew out of the authors' academic paper on extracting expert knowledge. Usually, when we want an LLM to respond in a specific manner, we write a system prompt of a few paragraphs. This works poorly: the model quickly falls back into standard politeness and forgets narrow-specialization technical constraints.

dot-skill solves the problem differently. The tool splits a person's profile into two independent layers:

1. The Persona layer handles tone, directness, mental models, and communication style.
2. The Work Skill layer stores architecture standards, typical workflows, response formats, and specific knowledge bases.

The resulting agent can not only repeat the former team lead's jokes, but also review API specifications exactly according to the rules that person followed during code review.

Three Usage Scenarios

Initially, the repository focused only on colleagues, but in recent versions the authors expanded the engine to three categories:

• Colleagues. The model trains on internal documents, pull requests, and work chats. You can invoke it entirely (/colleague-name) or access its technical part separately without personality quirks (/colleague-name-work).
• Close people (relationship). A mode for preserving the familiar communication style of friends or relatives based on message history. It captures triggers, dispute patterns, and reconciliation methods.
• Public figures (celebrity). Analyzes speeches, articles, commits, and interviews of famous engineers or thinkers, attempting to reproduce their decision-making logic rather than superficial style imitation.

Here's an interesting example from the repository, where the authors made a snapshot of Andrej Karpathy's thinking:

User            > Все говорят об агентах. Каким будет AI Agent в 2026 году?
celebrity.skill > Все пилят демки, но никто не занимается эвалюацией.
                  Узкое горлышко агентов — не интеллект модели, а скаффолдинг:
                  строгий тестовый стенд, пошаговый реплей ошибок,
                  контролируемый бюджет вызовов инструментов и четкий анализ сбоев.
                  В 2026 году я ставлю не на самую сильную модель,
                  а на лучший скаффолдинг и строгий eval.

Where to Get Data and How It Works

The quality of the resulting agent directly depends on the source material. If you feed the utility empty chatter from the break room, the result will be useless. The project creators state directly: detailed design documents, code comments, and reasoned decisions in trackers work best.

The package supports gathering information from multiple sources:

• Corporate chats (Slack via API, Feishu, DingTalk)
• Message exports and email dumps .eml / .mbox
• Local markdown files, notes, and PDF documents
• Subtitles from video presentations

For public figures, a separate pipeline is built into the repository:

# Скачиваем субтитры выступлений
bash tools/research/download_subtitles.sh "<video-url>" "./tmp/subtitles"

# Очищаем транскрипт и убираем таймкоды
python3 tools/research/srt_to_transcript.py "./tmp/subtitles/example.srt"

# Объединяем заметки по шести измерениям
python3 tools/research/merge_research.py "./skills/celebrity/<slug>"

# Проверяем структуру и качество артефактов
python3 tools/research/quality_check.py "./skills/celebrity/<slug>/SKILL.md"

Runtime Environment Support and Installation

The tool isn't tied to a single platform. Skills are generated as markdown manifests and run in most popular agentic environments: Claude Code, OpenClaw, Codex, DeepSeek Harness, and Hermes Agent.

Installing into an agent environment takes one line:

# Для ручной установки в Claude Code:
git clone https://github.com/titanwings/colleague-skill ~/.claude/skills/dot-skill

After that, the command /dot-skill becomes available in the agent's terminal. An interactive wizard will ask for the category, character name, key personality traits, and path to the source data.

If the bot behaves uncharacteristically for the prototype during conversation, you can correct it right in the dialogue with a phrase like "he wouldn't respond that way—he usually asks for the error trace immediately." The script intercepts the remark, records it in a correction layer, and instantly updates the instructions. A version manager with the ability to roll back to previous snapshots is provided for safe experimentation.

Is It Worth Trying

The project looks like a successful experiment at the intersection of practical automation and digital memory.

You'll find this useful if:

• You want to preserve the team lead's technical approach and standards before a team reorganization.
• You need a strict architecture reviewer who thinks in the categories of a specific technical director.
• You want to build a personal mentor from public materials of a famous engineer.

Two things are worth considering: privacy and data quality. Exporting corporate chat logs to unverified cloud models is a no-go for security reasons. But if you deploy a local pipeline and filter out secrets, dot-skill can turn scattered texts into a genuinely useful work assistant.