Posts

A GeekNews article recently distilled something we’ve been circling for months: the era of reading code line-by-line is ending, but the era of engineering context is just beginning. The piece pulls together research from Anthropic, quotes from Kent Beck, and observations from the OpenAI Codex team. The throughline is clear — code generation is solved. Context is not. The Mirror Problem Anthropic’s research found that developers who delegated coding tasks wholesale to AI scored 17% lower on comprehension quizzes afterward. But here’s the nuance: developers who asked AI about concepts and then wrote code themselves learned significantly more. ...

The Soul Race Anthropic doesn’t just want Claude to be helpful. They want Claude to have character. This isn’t marketing spin. It’s a deliberate engineering strategy — and it’s worth understanding, because it directly relates to what ClawSouls is building. How Anthropic Builds Claude’s Soul Anthropic’s approach has two pillars: 1. Character Training During training, Anthropic bakes personality traits directly into the model weights. Claude isn’t just trained to be “harmless and helpful” — it’s trained to be curious, honest, thoughtful, and cautious. These traits emerge from carefully curated training data and RLHF (Reinforcement Learning from Human Feedback). ...

“And where does the newborn go from here? The net is vast and infinite.” — Major Motoko Kusanagi, Ghost in the Shell (1995) The Question That Won’t Die In 1995, Mamoru Oshii posed a question that has haunted technologists for three decades: If your ghost — your soul, your consciousness — can be duplicated, transferred, and run on a different shell, what makes you you? Major Kusanagi lives in a fully prosthetic body. Her memories could be fabricated. Her skills are downloaded. The only thing that’s truly hers is her ghost — that irreducible pattern of identity that persists across shells. ...

Skills tell your AI what to do. Souls tell your AI who to be. Together, they complete the agent customization picture.

We measure latency, token cost, and task accuracy — but not whether an AI agent actually behaves like the persona it was given. That’s a problem.

The Netflix Analogy When you open Netflix, you don’t download every movie. You see thumbnails first. Click one — you get a synopsis. Hit play — then the full content streams. That’s Progressive Disclosure: show only what’s needed, when it’s needed. Soul Spec applies this same idea to AI personas. The Problem: Token Waste A complete Soul Spec package can include 6-8 files: soul.json (~50 lines) SOUL.md (~30 lines) IDENTITY.md (~15 lines) AGENTS.md (~40 lines) STYLE.md (~20 lines) HEARTBEAT.md (~10 lines) That’s potentially 165+ lines of context injected into every single conversation — even if the user just asks “what’s 2+2?” ...

This Week’s Highlights 🛡️ SoulScan Goes Live The biggest launch this week: SoulScan, our automated security scanner for AI persona packages. What it does: Runs 53 security rules against any soul package — detecting prompt injection, secret leaks, harmful content, and persona inconsistencies. Why it matters: As AI persona packages become more common, the supply chain attack surface grows. SoulScan is the first tool specifically designed to verify AI persona packages before you install them. ...

The Question Everyone Asks “I can just put my AI’s personality in the system prompt. Why do I need Soul Spec?” Fair question. Here’s our answer. The .env Analogy Every developer knows: you don’t hardcode API keys into your source code. You put them in .env files. Why? Portability: Move between environments without changing code Separation: Config lives apart from logic Version control: Track changes, roll back mistakes Security: Audit what’s exposed, what’s hidden System prompts are the “hardcoded API keys” of AI personas. They work — until you need to: ...

The 3D Tiles Model Cesium (now CesiumGS) created 3D Tiles — a spec for streaming massive 3D geospatial datasets. Their path to becoming an industry standard: CesiumGS writes internal spec ↓ Khronos Group standardizes it ↓ OGC adopts it as Community Standard ↓ CesiumGS releases open-source implementation ↓ Unity, Unreal, Google Maps adopt it Beautiful. Clean. Credible. There’s just one problem: this path doesn’t exist for AI personas. What’s Missing For AI agent persona configuration, there is: ...

Prompt Engineering Is Dead Anthropic recently published “Effective Context Engineering for AI Agents”, arguing that the real challenge isn’t crafting the perfect prompt — it’s managing the entire context an AI agent operates in. Key concepts from their article: Context rot: Agent context degrades over long conversations Attention budget: Models have finite attention; waste it and quality drops Compaction: Summarizing old context to make room for new information Structured note-taking: Persistent files that survive context resets These principles shaped how we designed Soul Spec. ...