🧠 Digital Stoic Praxis

Praxis (πρᾶξις) = practice. Knowledge enacted, not just known.

---

🎯 In One Sentence

A cognitive discipline for human-AI collaboration built on mutual sharpening — the human sets intent and directs, the AI challenges and reveals blind spots, and both learn from the loop.

⚠️ Live experiment. My cognitive toolkit — fork it, adapt it to your brain.

This is v2, and it’s smaller than v1. I built 105 skills, a 5-phase workflow, and a 9-skill gate system to force structure out of an AI that couldn’t yet plan its own work. Then the models got dramatically better at deciding and self-planning — and most of that scaffold became dead weight. Months of real practice wore it down to its handle. What survived isn’t a lighter version of the same thing; it’s what stays irreducibly human once the model can plan for itself: persist intent, guide at the big boundaries, challenge the confident answer. The arc — build it, practice it, shed it — is the praxis.

---

💡 Why This Exists

AI tools are powerful but chaotic. Most people either micromanage every prompt or let AI run wild and pray. Neither scales to real work.

This toolkit treats AI collaboration as cognitive discipline — a set of thinking modes you activate depending on the situation, where the AI also pushes back on your assumptions. The goal is not to make AI do more. It’s to think better together — and that includes letting AI challenge the human.

Deeper why → PHILOSOPHY.md

---

🧭 The Flow

The practice no longer fans out into phases. As the model got better at planning its own work, what’s left for the human collapsed to three things the model doesn’t do for itself:

flowchart TD
  M["🚀 Model self-plans(handles what scaffold used to force)"]
  subgraph CORE["what stays the human's job"]
    P1["💾 PERSISTcarry intent across sessions"]
    P2["🧭 GUIDEpropose, don't act ·gate at the big boundaries"]
    P3["😈 CHALLENGEhold the tension ·counter the confident model"]
  end
  M --> CORE
  CORE -->|"daily spine still runs underneath"| S["save/load-context · capture · commit · edit-toolkit"]
  classDef mod fill:#B3D9FF,stroke:#333,color:#000
  classDef pillar fill:#90EE90,stroke:#333,color:#000
  classDef spine fill:#E0E0E0,stroke:#333,color:#000
  class M mod
  class P1,P2,P3 pillar
  class S spine

You still frame, think, build, debug as needed — but they’re moves you reach for, not a lifecycle you march through. A model that self-plans doesn’t need the phases enforced. (/brainstorm, /troubleshoot, /frame-problem are still here when a problem is genuinely hard — see the full catalog.)

Plus: tool creation (/edit-tool), conversions (PDF, EPUB, Google Docs), and domain plugins (GTD, coaching, business analysis, philosopher personas).

---

🤝 Human-AI Governance

Not human-drives-AI. Not AI-drives-human. Mutual sharpening:

• Human → AI: Sets intent, chooses workflow phase, directs execution, verifies at gates
• AI → Human: Challenges assumptions, surfaces blind spots, reveals patterns the human can’t self-observe
• Together: The human thinks more clearly because the AI pushes back. The AI produces better work because the human sets precise intent.

Key: Posture before technique — but posture includes the willingness to be challenged.

Full autonomy spectrum + orchestrated agency → PRACTICE.md

---

💰 Cognitive ROI (Return on Tokens)

Not “how fast” — how deep:

Maturity = shifting tokens from ⚙️ toward 🧠.

Multiplied by context efficiency (don’t waste tokens re-explaining) and compounding (learnings reduce future spend). Full model → PRACTICE.md

---

📊 By the Numbers

• 128 skills across 16 plugins — but only ~12 run daily; the rest are parachutes (rare by design) or an archive of exploration. Lead with the spine, not the count.
• 16 plugins: core (dstoic), cognitive, openspec, content, convert, toolsmith, experimental, retrospect, GTD, coaching, business analysis, philosopher personas, cowork, lazy, team, ring
• 2 agents: devil’s advocate, context summarizer
• Hooks: 4 live (notify-tmux, dump-output, retrospect-capture, check-praxis-dir) + experimental staging
• 3 execution modes: garage (default), scale, maintenance
• 🪥 Toothbrush principle: This is one practitioner’s discipline. Don’t copy — adapt. Why?
• 📊 Benchmarked against 7 frameworks including ECC (144K⭐), ACP, BMAD (36K⭐) — leads on context quality, cognitive depth, adversarial thinking. Only BMAD edges ahead (for teams). Details → PRACTICE.md · /benchmark-praxis skill

---

🏪 Plugins

📦 Install

/install-plugin https://github.com/digital-stoic-org/agent-skills

To install a specific plugin only, add it to .claude/settings.json:

{"plugins": ["digital-stoic-org/agent-skills/dstoic"]}

🚀 Quick Start

/frame-problem how should I approach building a new auth system
# → Triangulates: Keogh=2 (team expertise), Predictable=yes, Disassemble=yes
# → Complicated (3/3 agree) + Boulder → /investigate → /openspec-plan

/brainstorm product naming ideas for my CLI tool
# → Research → SCAMPER divergence → weighted scoring → recommendation

/troubleshoot "TypeError: Cannot read property 'map' of undefined"
# → WebSearch → qualify → diagnose → OODA → save learning

---

📖 Documentation — Two Cognitions, One Practice

This toolkit serves two kinds of readers with fundamentally different needs:

👤 Human cognition — scans, skims, needs cognitive ease. Overwhelmed by walls of text. Seeks the “aha” moment, then dives deeper only when motivated. Diagrams and structure reduce cognitive load.

🤖 LLM cognition — parses tokens. Needs directive density, structured data, zero filler. Diagrams are nice but YAML is better. Self-contained is better than linked.

The practice itself is about combining both: a human who thinks consciously and an AI that both amplifies and challenges that thinking. The documentation mirrors this — each layer optimized for its reader:

Human docs link to each other (progressive depth). LLM docs are self-contained (no navigation, everything in one shot). The practice works because both cognitions play to their strengths.

---

⚠️ 🪥 CLAUDE.md = toothbrush. See CLAUDE.md.example for inspiration, don’t copy. (Why?)

The example uses rtk for token-optimized command output. Install it separately if you want to use the rtk instructions.

💡 IDEAS.md — Future ideas & what I’m deliberately not building yet.