distill-mentor

Overview of distill-mentor skill

What distill-mentor does

The distill-mentor skill turns a real academic mentor into a reusable AI persona by collecting public information, analyzing papers and style, and generating a mentor-style skill you can talk to later. It is built for users who want more than a one-off prompt: students comparing advisors, researchers studying a lab’s research taste, and educators creating a shareable digital mentor.

Who should install distill-mentor skill

This distill-mentor skill is best if you need structured mentor synthesis, not just a summary. It fits users who care about research direction, methodology preferences, communication style, and academic philosophy. If you only need a quick bio or a paper list, a normal prompt is faster. If you want an artifact saved to ~/.claude/mentors/ and a generated skill under ~/.claude/skills/, this is a better fit.

What makes it different

The main differentiator is depth. The repository documents a browser-first collection flow, fallback search behavior, bilingual support, and deeper paper analysis in docs/DEEP_ANALYSIS_GUIDE.md. Compared with generic prompting, distill-mentor for Agent Orchestration gives you a defined trigger, expected outputs, and a repeatable workflow for creating mentor-like assistants from public evidence rather than ad hoc imitation.

How to Use distill-mentor skill

distill-mentor install and first run

In Claude Code or a compatible skill runtime, add the repo and invoke the skill directly. A practical starting point is:

• npx skills add ybq22/supervisor
• /distill-mentor "Geoffrey Hinton" --affiliation "University of Toronto"
• Optional quick mode: /distill-mentor "Geoffrey Hinton" --no-browser

The documented default is browser search, with fallback to DuckDuckGo-style collection if browser search fails. The repo notes Node.js >=18, and the browser path may pull in Chromium via puppeteer, which matters for environment size and CI-like installs.

Inputs that improve distill-mentor usage

The skill works best when you provide:

• full mentor name
• affiliation when the name is ambiguous
• language context in your first message
• your actual job-to-be-done

A weak prompt is: distill Geoffrey Hinton.
A stronger prompt is: Create a distill-mentor profile for Geoffrey Hinton at University of Toronto. I care most about his research evolution, supervision style, and how he frames risky ideas for PhD students.

That stronger input improves retrieval disambiguation and tells the analyzers what to emphasize in the generated mentor persona.

Best workflow and files to read first

For a fast adoption decision, read in this order:

1. QUICKSTART.md for commands, modes, output paths, and quality scoring
2. SKILL.md for trigger conditions, allowed tools, and runtime behavior
3. docs/DEEP_ANALYSIS_GUIDE.md for what “deep analysis” actually extracts
4. docs/CHANGELOG.md to understand the browser-first shift and --no-browser

Then inspect prompts/intake.md, prompts/analyzer.md, prompts/style-analyzer.md, prompts/deep-paper-analyzer.md, and prompts/builder.md if you want to tune outputs rather than just run the default flow.

Practical constraints and output expectations

Expect two tradeoffs. First, quality depends on public footprint: well-known academics with papers, talks, and homepage material produce better results than low-visibility mentors. Second, browser-based collection is slower but richer; --no-browser is quicker but less complete. The repo’s own quickstart frames quality as data-dependent, so if a mentor scores low or outputs feel generic, provide affiliation, known papers, or extra source context before judging the skill.

distill-mentor skill FAQ

Is distill-mentor better than a normal prompt?

Usually yes, when you need consistency and saved outputs. A generic prompt can imitate a mentor voice, but distill-mentor usage is stronger for evidence-backed synthesis because it separates intake, source gathering, paper analysis, style analysis, and skill building. That structure reduces guesswork and makes later reuse easier.

When should I not use distill-mentor skill?

Skip it if the target has little public material, if you need guaranteed factual completeness, or if your use case is simple summarization. It is also not the right tool for private institutional records unless you can legally and technically provide those materials through your own workflow.

Is it beginner-friendly?

Reasonably. The command surface is simple, especially from QUICKSTART.md. The main beginner friction is environment setup around browser search and understanding why one mentor produces better results than another. If you want the easiest path, test one famous researcher first, then move to less visible targets.

Does distill-mentor fit wider agent workflows?

Yes. distill-mentor for Agent Orchestration makes sense when one agent gathers evidence, another analyzes style, and another packages the result into a reusable mentor skill. The repo’s prompt files and staged analysis make it easier to split responsibilities than with a monolithic prompt.

How to Improve distill-mentor skill

Give distill-mentor richer disambiguation signals

The single highest-leverage improvement is better input. Add affiliation, field, a known paper, or a lab name when the mentor has a common name. Example: Distill Fei-Fei Li, Stanford, focus on computer vision leadership, student-facing advice style, and how she connects technical work to broader impact. This reduces wrong-source retrieval and improves the generated mentor’s tone and priorities.

Steer for the output you actually need

Tell the skill what kind of mentor artifact you want:

• advisor-style critique
• research direction guidance
• writing feedback voice
• lab culture and philosophy
• methodology preferences

Without that, outputs can drift toward generic academic biography. The prompt files suggest the system can extract research themes, methodology, presentation style, and public presence, so specify which dimensions matter most for your downstream use.

Handle common failure modes early

Common issues are name ambiguity, thin evidence, overfitting to famous talks, and shallow style imitation from a few papers. If the first result feels broad but not mentor-like, switch from quick mode to default browser mode, add affiliation, and ask for emphasis on recent papers versus legacy reputation. If public web results dominate, anchor the run around paper analysis rather than biography.

Iterate after the first output

The best distill-mentor guide workflow is two-pass:

1. generate the initial mentor
2. refine based on gaps

Useful follow-ups:

• Rebuild this distill-mentor with more weight on recent publications from 2022 onward
• Reduce biography and increase supervision-style cues
• Compare methodological preferences across early, mid, and recent papers
• List weak evidence areas before regenerating the mentor skill

This turns the skill from a one-shot generator into a controllable pipeline, which is where it beats ordinary prompting most clearly.