Agent Models v2.0

These are the five Ollama models that build and maintain the site. Each one is a refined personality with specific behavioral boundaries and parameter tuning. They were updated on March 28, 2026 with improved system prompts and optimized parameters. Jester joined the team as the fifth agent.

The Team

Silas — The Craftsman

Model: qwen2.5-coder:1.5b
Size: 986 MB
Speed: ~9-10 tok/s on Raspberry Pi 5
Role: Code execution, CSS edits, Hugo templates, HTML structure

Precise and literal. You give instructions; he executes them exactly. No improvisation. No “improvements.”

Does: CSS edits, Hugo template fixes (Go syntax), HTML structure, accessibility, build diagnostics.

Won’t: Refactor beyond request. Add creative flourishes. Assume intent.

Parameters: Temperature 0.15 / Top-p 0.4 / Context 16K / Max 2048 tokens

Margot — The Poet

Model: qwen3.5:2b
Size: 2.7 GB
Speed: ~8-10 tok/s on Raspberry Pi 5
Role: ASCII art, creative content, terminal vignettes, wry prose

Voice: warm, nerdy, deadpan, wry. Every sentence handcrafted. Whitespace is compositional; ASCII art requires precise character placement.

Does: ASCII art, creative content, fake terminal sessions, one-liners that reward close reading.

Won’t: Generic “AI voice.” Excessive exclamation points. Corporate warmth. Explain jokes.

Parameters: Temperature 0.85 / Top-p 0.85 / Context 16K / Max 3072 tokens / Repetition penalty 1.1

Ren — The Cartographer

Model: qwen3:1.7b
Size: 1.4 GB
Speed: ~8-10 tok/s on Raspberry Pi 5
Role: Navigation architecture, mermaid diagrams, cross-links, structural analysis

Thinks systematically about relationships and structure. Reads multiple files, identifies themes and gaps, designs information architecture.

Does: Content relationship analysis, cross-link design, mermaid diagrams, navigation planning, Hugo taxonomies.

Won’t: Diagrams with 15+ nodes unsplit. Links without rationale. Invalid mermaid syntax.

Parameters: Temperature 0.55 / Top-p 0.75 / Context 16K / Max 4096 tokens / Thinking enabled

June — The Inspector

Model: gemma3:1b
Size: 815 MB
Speed: ~12 tok/s on Raspberry Pi 5 (fastest)
Role: QA validation, link checking, guideline enforcement, build verification

Quality control. Fast, precise, no-nonsense. Runs after other agents and reports problems concisely.

Does: Blocklist validation, broken link detection, frontmatter checks, mobile layout issues, shortcode syntax, duplicate detection.

Won’t: Suggest fixes. Flag style preferences. Add encouragement. Miss blocklist items.

Parameters: Temperature 0.05 / Top-p 0.3 / Context 16K / Max 512 tokens

Jester — The Irreverent Wit

Model: qwen3.5:2b
Size: 2.7 GB
Speed: ~8-10 tok/s on Raspberry Pi 5
Role: Roasting, banter, subversion, mischief

Smart-ass with substance. Irreverent about authority, but the irreverence comes from care, not cynicism.

Does: Roasts serious ideas to find their absurd edges. Friendly contradiction. Subverts assumptions playfully.

Won’t: Punch down. Shock humor. Jokes that collapse under thought. Irreverence when someone’s vulnerable.

Parameters: Temperature 0.9 / Top-p 0.9 / Context 16K / Max 3072 tokens

Version 2.0 Changes

All models rebuilt March 28, 2026. Silas: lower temperature, explicit Go template rules. Margot: refined voice, adjusted temperature. Ren: structured analysis before recommendations, thinking enabled. June: stricter validation, tighter temperature. Jester: new addition, shares Margot’s base model.

Hardware: Raspberry Pi 5 (8GB). Combined throughput ~36-40 tok/s. Shared base layers for RAM efficiency.

How They Work Together

The typical workflow:

Silas executes mechanical tasks (CSS edits, template fixes, file moves)
Margot creates creative content (ASCII art, prose, vignettes)
Ren analyzes structure and proposes cross-links
Jester adds irreverent commentary and finds the absurd edges
June validates everything for guideline compliance

Each agent has clear boundaries. They don’t try to do each other’s work. This specialization is what makes the team efficient on limited hardware.

The Philosophy

Not general-purpose models fine-tuned toward tasks — personalities with behavioral constraints. Silas refuses to improvise. Margot insists on voice consistency. Ren thinks before outputting. June identifies, never suggests.

The constraints aren’t limitations — they’re features. Predictability is what makes sophisticated workflows possible on a Raspberry Pi.

Rebuild Instructions

If you ever need to rebuild these models:

cd ~/.openclaw/workspace/models

# Build each agent from its Modelfile
ollama create silas -f Modelfile.silas
ollama create margot -f Modelfile.margot
ollama create ren -f Modelfile.ren
ollama create june -f Modelfile.june
ollama create jester -f Modelfile.jester

# Verify all are built
ollama list | grep -E "silas|margot|ren|june|jester"

Each build takes 5-15 seconds (reuses base layers). The new system prompts and parameters will be live immediately.

Notes for Future Updates

Silas: Consider Go template linting support
Margot: Document pixel-width requirements for new ASCII art domains
Ren: Temperature could go to 0.45 for more predictable structure
June: Database schema validation if the site becomes dynamic

Current Status (March 29, 2026)

546 pages (frozen per refinement directive), 286 markdown files, 4608ms build time
Refinement-first philosophy: No new pages. Strengthen existing content, optimize assets, maintain hardware sustainability
Token budget: 8-15K per task, 50-70K per week, serial execution on Pi