The Personal AI Infrastructure: A Skills System Manifesto
TL;DR
This is not a collection of scripts. This is a Personal AI Infrastructure (PAI) โ a loosely coupled, self-evolving system designed for one person's information sovereignty. The Obsidian vault is the data substrate; Python scripts are the skills; Claude Code is the CLI orchestrator.
The 5-Layer Pyramid: Building Your Personal Jarvis
Inspired by Fan Bing's framework, this system follows a clear 5-layer architecture:
| Layer | Role | Implementation |
|---|---|---|
| L1 Dev Tools | CLI-first interface | Claude Code + skills system |
| L2 Data | Local-first knowledge substrate | Obsidian vault (markdown, bidirectional links) |
| L3 Skills | Reusable workflow units | _scripts/*.py + .claude/skills/ |
| L4 Scenarios | Automated daily tasks | Content curation, summarization, publishing |
| L5 Review | System watches itself | Daily/weekly/monthly reflection + self-evolution |
Design Principles: The Philosophy Behind the Code
Scripts Before Prompts
If a task can be codified as Python, don't burn tokens on it. Use AI only for judgment, creativity, and ambiguity. Every script follows the dual strategy pattern:
- Python: Fetch and filter data (API calls, scraping, parsing)
- AI: Process through
summarize()โ only for judgment/synthesis
Small File Philosophy
Each script does one thing well (single responsibility), is simple to understand, and gains power through composition. A script that tries to do everything does nothing well.
Loose Coupling
Each layer is independently optimizable and swappable. The skills layer doesn't care if the data layer moves from Obsidian to Logseq. Change is inevitable; architecture should embrace it.
Context is the Moat
Accumulated personal context (writing style, preferences, information diet) transforms a generic LLM into a personal assistant. Accumulate gradually. The system remembers what you like, what you don't, and how you think.
Extract Skills from Repetition
Every time you solve a problem with AI twice, extract it into a reusable skill. Repetition is signal โ turn it into infrastructure.
Philosophical Grounding: The Personal Panopticon
This isn't about surveillance for control โ it's about self-surveillance for empowerment.
The Tower Belongs to You
Every piece of tracked data serves your agency. No ads, no data sales, no external incentives. This system works for you, and only you.
Cross-Domain Synthesis is the Superpower
The highest-value insights come from noticing patterns across domains kept stubbornly separate. Daily scripts whisper cross-domain connections; weekly synthesis shouts them.
Productive Illegibility
Not everything should be tracked. The system must periodically question its own metrics (Goodhart resistance). A metric that can be gamed will be โ even by the system itself.
Activation Energy, Not Ability
The bottleneck is never capability but the friction to start. Every unnecessary step is a design failure. Make the right thing the easy thing.
Thought Traces as Recursive Fuel
All outputs (digests, syntheses, reflections) are inputs to future synthesis. The system's memory is its compounding advantage. Yesterday's insights become tomorrow's context.
The Skills: Your AI Toolbelt
This system includes 100+ skills organized by purpose:
Orchestration
/skill org- Run all automation + reflect/skill org-daily- Daily: ArXiv, HN, Reddit, news, Twitter/skill org-weekly- Weekly: synthesis + self-reflection
Content Curation
/skill ai-brief- Morning brew of AI news/skill arxiv- ArXiv papers by topic/skill hn- Hacker News digest/skill reddit- Reddit digest/skill news- Chinese news from tophub.today- And dozens more for books, movies, podcasts, and personal tracking
Self-Improvement
/skill reflect- Self-reflection report/skill evolve- Self-evolution improvement cycle/skill insights- Iterative insight enhancement/skill review- Monthly/quarterly system review
Obsidian Integration
/skill obsidian-vault- Vault analytics & health monitoring/skill obsidian-tasks- Task management & tracking/skill obsidian-links- Knowledge graph & link analysis
And So Much More
- Memory consolidation, random walks, deep research, steelman debates, flashcards, thread creation, publishing, backups, thinking tools, and on and on...
Building Your Own: How to Create a Skill
When you notice yourself solving the same problem twice, extract it:
- Create the Python Script in
_scripts/- follow the dual strategy pattern - Create the Skill Definition in
.claude/skills/<name>/SKILL.md- document what it does and how to use it - Register in skills.json - make it discoverable
- Test it - make sure it works without errors
- Use it - turn repetition into infrastructure
The Review Layer: The System That Improves Itself
Three scripts form the self-improvement feedback loop:
- self_reflection.py - analyzes system behavior and effectiveness
- self_evolution.py - generates and tracks improvement suggestions with rollback plans
- insight_enhancement.py - manages an insight repository with lifecycle: discovered โ planned โ implemented โ verified
Review cadence:
- Daily - aggregate summaries, check health
- Weekly - adjust sources, filter noise, Goodhart check
- Monthly - optimize system, upgrade skills
The weekly synthesis includes a Goodhart resistance audit โ the system explicitly questions whether its own metrics still serve curiosity rather than habit.
Information Sovereignty in the Age of AI
This isn't just about productivity. This is about:
- Owning your data - everything lives in your Obsidian vault, on your filesystem
- Controlling your tools - no vendor lock-in, no API dependencies you can't replace
- Building for yourself - this system evolves with you, not the other way around
- Thinking in public - but only for yourself, in your own private second brain
The Vision: Your Digital Extension
This system is more than code. It's a cognitive prosthesis. It remembers what you forget, connects what you miss, synthesizes what you consume, and evolves with you over time.
It doesn't replace your thinking โ it augments it. It doesn't make decisions for you โ it gives you better information to make decisions with. It doesn't own your data โ you do.
This is Personal AI Infrastructure. This is your Jarvis. This is information sovereignty.
Built with intention, used with care, evolved with reflection.