Refine 'Tools as Primitives' principle with spectrum-based guidance

[kieranklaassen]

Summary

Addresses #71 - This PR refines the agent-native-architecture skill's guidance on tool design, replacing the binary "primitives vs workflows" stance with a more nuanced spectrum-based approach.

Key Changes

New Sections Added

1. Primitives Vocabulary - Defines key terms upfront:

   • Pure Primitive, Guided Primitive, Domain Tool, Orchestrated Action, Sub-Agent, Harness
   • Clarifies that a "tool call" is just a mechanism—what you wire to it determines its nature

2. "What IS a Tool?" - Philosophical section covering:

   • API doesn't distinguish between file reads and sub-agent spawns
   • Sub-agents break the primitive model (they make their own decisions)
   • Tool descriptions as JIT context injection

3. Decision Framework - New mental model:

   Decision Owner	Implementation
   Business rule (always this way)	Code in tool execution
   Tool-specific strategy (usually this way)	Tool description
   Conversation-dependent	System prompt
   Agent judgment	Neither—trust the model

4. Harness Pattern - When to use code to enforce business invariants

Updated Sections

1. Granularity Principle - Now shows the spectrum:

   Pure Primitive → Guided Primitive → Domain Tool → Orchestrated Action
   

   With guidance on when to use each.

2. Anti-Patterns - Restructured with:

   • "When it's wrong" vs "When it's fine" for each pattern
   • Acknowledges that logic in tools isn't automatically wrong
   • Separates "judgment calls" from "actual anti-patterns"

3. Architecture Checklist - Updated with new checkboxes:

   • Right level of granularity
   • Decision ownership clear
   • JIT Context Injection

Closes

Closes #71