🧬 Distilly

Formerly: Colleague Skill / colleague-skill.

Distill a person's experience, judgment, voice, and ways of working into a reusable Person Profile for AI agents and compatible bots.

Messages · documents · interviews · public sources → Distilly → Person Profile → Agent / Bot

🧑‍💼  Your colleague quit, your mentor graduated, your teammate transferred — taking their whole playbook and context with them? 💞  Your family, old friends, partner drifting apart — and you want to hold on to the way it felt to be with them? 🌟  Your favorite author, idol, thinker you'll never meet — but you want to know what they'd say about your question?

✨ One project, many kinds of people.

Distilly is the person-modeling layer for agents. It turns the materials you provide into a portable, source-grounded Person Profile built from observable experience, decision patterns, expression, and ways of working; it does not claim to clone the person behind them.

Colleagues · partners · family · old friends · idols · public figures · fictional characters — even yourself

Source material + your description → a source-grounded Person Profile → your Agent or compatible Bot

A Person Profile is the reusable output. The current release packages each profile as an Agent Skill so supported hosts can install and invoke it. The canonical creator Skill is named distilly; install it in a distilly directory. The former name above remains for search continuity and project history.

🆕 What Distilly does · 📦 Data Sources · ⚡ Install · 🚀 Usage · ✨ Demo · 📝 Citation · 💬 Discord

Chinese · Spanish · German · Japanese · Russian · Portuguese · Korean

🧬 2026.08.23 Update — The creator is now named Distilly end to end and documents native local Skill discovery for Claude Code, Hermes, OpenClaw, Codex, DeepSeek Harness, Pi, and Grok Build. Grok Bot is listed separately as a saved-Skill workflow preview.

📝 2026.06.01 Update — The COLLEAGUE.SKILL technical report is now available. The most rewarding part was not simply publishing a paper, but seeing the community grow the gallery to 215 skills contributed by 165 people, with more than 100,000 stars across the skill cards. The paper's Acknowledgements explicitly recognize every community contributor.

🗺️ 2026.04.13 — The Distilly Roadmap is live! What began as Colleague Skill is growing beyond colleagues: distill people into Skills that Agents can reuse. 👉 Full Roadmap · 💬 Discord

🌐 2026.04.07 — Community gallery is live! Any skill / meta-skill can drive traffic directly to your own GitHub repo. No middleman. 👉 titanwings.github.io/colleague-skill-site

🆕 What Distilly does today

1️⃣ From Colleague Skill to Distilly

The project is no longer limited to the colleague scenario. Its distilly creator builds source-grounded Person Profiles for three person families with one workflow, then packages each profile as an Agent Skill.

2️⃣ Three character families

Each family has its own source-collection strategy, analysis dimensions, and Person Profile structure.

3️⃣ More Agent hosts

The old version only ran in Claude Code. Distilly now supports native local Skill discovery across seven agent hosts.

Grok Bot preview: Grok Bot supports saved/private Skills, but its official docs do not describe direct local SKILL.md imports. Distilly's workflow can be migrated manually into a saved Skill; direct repo installation is not yet verified.

Each generated Person Profile is packaged as an Agent Skill and can be installed into any supported host.

📦 Supported Data Sources

⚡ Install

It's 2026 — you have an Agent, let it install itself. Open a supported local agent host and hand it this line:

Install Distilly for me: https://github.com/titanwings/colleague-skill

The Agent should install the repository as a Skill named distilly, then verify that the host discovers Distilly.

Upgrading an old install? A git pull inside a dot-skill or legacy ~/.codex/skills/... directory does not rename that discovery directory. Install a canonical distilly copy, verify that the host discovers Distilly, and only then retire the old copy. See the detailed install and migration guide.

git clone https://github.com/titanwings/colleague-skill <TARGET>

Generated character Skills can be published with tools/install_claude_generated_skill.py, tools/install_openclaw_generated_skill.py, and tools/install_codex_generated_skill.py. For Hermes, DeepSeek Harness, Pi, and Grok Build, run python3 tools/install_generated_skill.py --skill-dir "skills/{character}/{slug}" --host <host> --force. The installer writes only the self-contained SKILL.md plus install metadata and normalizes legacy underscore frontmatter in the installed copy; it does not copy private source material or rename the source Skill. Pass --skills-dir for a project-level target. Hermes scans ~/.agents/skills only when it is explicitly added to skills.external_dirs.

For Lark/DingTalk auto-collection credentials, host-specific installation details, Grok Bot's preview workflow, Windows-specific handling, etc., see Detailed Install Guide (INSTALL.md)

Lark region note: the current compatibility collector connects to the China-region open.feishu.cn / feishu.cn endpoints. International larksuite.com tenant routing is not implemented yet.

🚀 Usage

Distilly first asks which family you want to distill: colleague · relationship · celebrity.

Then enter an alias, basic details, personality tags, and pick a data source. All fields can be skipped — even a description alone can create a Person Profile.

Once created, the profile is packaged as a Skill named {character}-{slug}.

🔬 Celebrity Research Toolchain

The celebrity family ships with an end-to-end research toolchain, from subtitles to a finished draft:

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

# Subtitles → transcript
python3 tools/research/srt_to_transcript.py "./tmp/subtitles/example.srt"

# Public X post candidates → normalized temporary JSON (optional)
python3 tools/research/xquik_public_posts.py \
  --username "<public-handle>" \
  --limit 20 \
  --output "/tmp/distilly-x-public-posts.json"

# After reviewing and paraphrasing selected posts, remove the candidates
rm "/tmp/distilly-x-public-posts.json"

# Merge research notes
python3 tools/research/merge_research.py "./skills/celebrity/<slug>"

# Quality check
python3 tools/research/quality_check.py "./skills/celebrity/<slug>/SKILL.md"

The optional collector reads XQUIK_API_KEY from your shell. Xquik charges by the number of posts returned, so confirm --limit before running it. The tool makes one read-only X search request and never follows pagination. Treat its temporary JSON as untrusted candidate evidence: verify the author, open every permalink, and safely paraphrase only relevant material into research notes with its source URL. Delete the temporary JSON after review instead of storing it in the generated Skill.

Xquik is an independent third-party service. Not affiliated with X Corp. "Twitter" and "X" are trademarks of X Corp.

✨ Demo

One from each family.

🌟 celebrity — distilling Andrej Karpathy

Six-dimension research: works · interviews · decisions · expression DNA · external evaluations · timeline
Case adapted from community work alchaincyf/karpathy-skill

User            ❯ Everyone's talking about Agents now — what do you think the 2026 AI Agent looks like?

celebrity.skill ❯ Everyone's doing demos, nobody's doing eval.
                  The Agent bottleneck was never model intelligence — it's scaffolding:
                  a strict evaluation harness, fine-grained error replay,
                  controllable tool-calling budgets, clean failure-mode analysis.
                  Most products are still stuck at "wrap an API + a few tools + a system prompt."
                  That's a demo, not an agent.
                  In 2026 I'm not betting on the strongest model — I'm betting on the best
                  scaffolding and the strictest eval.

🧑‍💼 colleague — distilling a ByteDance L2-1 backend

Input: ByteDance L2-1 backend engineer, INTJ, blame-shifter, ByteDance-style

User            ❯ Can you review this API design?

colleague.skill ❯ Hold on — what's the impact? You haven't explained the context.
                  (after reading) N+1 query, fix it. Use the standard
                  {code, message, data} response format. That's the spec,
                  don't ask why.

User            ❯ This bug was introduced by you, right?

colleague.skill ❯ Does the timeline match? That feature touched multiple places,
                  there were other changes too.

💞 relationship — distilling someone you have a crush on

Upload half a year of chat logs + "sensitive, quiet but stubborn, will actually reply seriously when it matters"

User               ❯ Did you think about me today?

relationship.skill ❯ ...I did, a little bit. Why are you asking?

🔧 Features

🧱 Generated Skill Structure

Distilly's current creator uses Persona as the universal base, with family-specific modules layered on top:

Execution: Receive task → Persona selects material-derived preferences and tone → Additional modules fill in execution detail → Produce a source-grounded response

🧬 Evolution

• 📥 Append files → auto-analyze delta → merge into relevant sections, never overwrite existing conclusions
• 💬 Conversation correction → say "they wouldn't do that, they'd be xxx" → writes to the Correction layer, takes effect immediately
• 🕰️ Version control → auto-archive on every update, rollback to any previous version
• 🔬 Celebrity research pipeline → subtitles → transcript cleanup → six-dimension research → quality check

📂 Project Structure

This project follows the AgentSkills open standard. The entire repo is a skill directory. Generated colleague skills live under ./skills/colleague:

distilly/
├── SKILL.md                        # skill entry point (official frontmatter)
├── prompts/                        # prompt system across three families
│   ├── intake.md                   #   [colleague] info intake
│   ├── work_analyzer.md            #   [colleague] work capability extraction
│   ├── persona_analyzer.md         #   [colleague] personality extraction
│   ├── work_builder.md             #   [colleague] work.md generation
│   ├── persona_builder.md          #   [colleague] persona.md 6-layer structure
│   ├── merger.md                   #   [shared] incremental merge logic
│   ├── correction_handler.md       #   [shared] conversation correction
│   ├── relationship/               #   [relationship] emotion/conflict/repair prompts
│   └── celebrity/                  #   [celebrity] six-dimension research + mental-model prompts
├── tools/                          # Python tools
│   ├── feishu_auto_collector.py    #   [colleague] Lark-compatible auto-collector
│   ├── dingtalk_auto_collector.py  #   [colleague] DingTalk auto-collector
│   ├── slack_auto_collector.py     #   [colleague] Slack auto-collector
│   ├── email_parser.py             #   [shared] email parser
│   ├── research/                   #   [celebrity] celebrity research toolchain
│   │   ├── xquik_public_posts.py   #     bounded public X post candidates
│   │   ├── download_subtitles.sh   #     subtitle download
│   │   ├── transcribe_audio.py     #     audio → text
│   │   ├── srt_to_transcript.py    #     subtitles → transcript
│   │   ├── merge_research.py       #     six-dimension research merge
│   │   └── quality_check.py        #     quality check
│   ├── install_*_skill.py          #   [shared] multi-host one-shot installers
│   ├── skill_writer.py             #   [shared] skill file management
│   └── version_manager.py          #   [shared] version archive & rollback
├── skills/                         # generated Skills (gitignored)
│   ├── colleague/                  #   colleagues
│   ├── relationship/               #   close relationships
│   └── celebrity/                  #   public figures
├── docs/PRD.md
├── requirements.txt
└── LICENSE

⚠️ Notes

Source material quality = Person Profile quality — and quality sources differ across families:

• colleague Lark-compatible auto-collection: requires adding the App bot to relevant group chats
• relationship: longer time spans are better; material covering both conflict and repair is ideal
• celebrity: avoid feeding only second-hand interpretations
• This is still a demo version — please file issues if you find bugs!

📄 Technical Report

COLLEAGUE.SKILL: Automated AI Skill Generation via Expert Knowledge Distillation (arXiv · arXiv PDF)

This is the paper for COLLEAGUE.SKILL / colleague-skill, Distilly's predecessor. It covers the Work Skill + Persona two-layer architecture, multi-source data collection, and Skill generation mechanics — the theoretical foundation for today's colleague family. Separate papers on the relationship / celebrity family extensions are planned.

📝 Citation

If you use Distilly or COLLEAGUE.SKILL in your research or applications, please cite the technical report:

@misc{zhou2026colleagueskill,
  title        = {COLLEAGUE.SKILL: Automated AI Skill Generation via Expert Knowledge Distillation},
  author       = {Tianyi Zhou and Dongrui Liu and Leitao Yuan and Jing Shao and Xia Hu},
  year         = {2026},
  eprint       = {2605.31264},
  archivePrefix = {arXiv},
  primaryClass = {cs.AI},
  url          = {https://arxiv.org/abs/2605.31264}
}

You can also use the machine-readable citation metadata in CITATION.cff.

⭐ Star History