Divinci Council

"Like a good council, Divinci is there... generating content at 2 AM while you sleep." — Sir Spamalot

A multi-persona deliberation pipeline that transforms blog posts into platform-native content and reviews code architecture and generates comedic scripts about AI industry mishaps. Runs on local Ollama, cloud APIs, or your current Hermes session.

Think of it as a writers' room that happens to include a Renaissance inventor, an Atlantean archivist, a viral growth hacker, a LinkedIn thought leader, a VEO director, a Monty Python parodist, a brand guardian, and now — a code review cursor.

SKILL.md · v3.0 Release Notes · The Show

---

Quick Start

# Clone into your Hermes skills directory
git clone https://github.com/Divinci-AI/divinci-council.git \
  ~/.hermes/skills/creative/divinci-council

# Or run standalone
cd divinci-council
cp templates/council.yaml ./council.yaml
cp templates/brand-guide.md ./brand-guide.md
# Edit council.yaml with your project settings
python3 scripts/orchestrator.py --batch 001

Prerequisites: Python 3.10+, PyYAML, and at least one backend:

• Ollama running locally (recommended for bulk)
• OpenCode CLI for cloud models
• cursor-agent for code review
• Or just use Hermes native — no extra installs

---

What It Does

The council reads your blog posts and generates:

Output	Count	Personas
Tweets	7-15	Leonardo, Archivist, Gemma4, Growth Hacker, Sir Spamalot
LinkedIn Posts	2	Thought Leader
VEO3.1 Scripts	3	VEO Director, Sir Spamalot (parody)
Show Scenes	∞	All 9 (diegetic content for "The Release Cycle")

And now in v3.0:

Output	Trigger	Personas
Code Review	--code-review file.py	The Cursor
Architecture Review	--architecture-review	The Cursor + Leonardo + Thought Leader
Meta-Review	--meta-review	Brand Guardian synthesizes all 9 lenses
Self-Reflection	After N batches	Council reviews its own outputs and methodology

Every piece of content is reviewed by the Brand Guardian for forbidden words, tone consistency, and thematic motifs.

---

The Nine Council Members

Member	Voice	Specialty
Leonardo	Renaissance inventor sketching flying machines	Conceptual hooks, architectural insights
Atlantean Archivist	Keeper of crystal-memory, speaks in harmonics	Poetic-prophetic framing, lower-case flow
Gemma4	Earnest junior engineer, Star Wars references	Open-weights perspective, edge-case hunter
Viral Growth Hacker	Engagement-obsessed X native	Algorithm-aware virality optimization
LinkedIn Thought Leader	Contrarian-but-helpful executive	Data-dense long-form authority posts
VEO3.1 Director	Thinks in shots, beats, visual gags	Timestamped video scripts, no lip-sync
Sir Spamalot	Monty Python + State Farm parody genius	Absurdist, legally-safe transformative parody
Brand Guardian	Sober final check	Quality gates, forbidden-word enforcement, security
The Cursor (NEW in v3.0)	Precision instrument, diff-native	Code review, refactoring, architecture deliberation

Each member can be routed to a different model backend. Leonardo runs on your laptop via Ollama. The Cursor delegates to cursor-agent CLI. The Brand Guardian uses whatever deep-reasoning model your Hermes session provides.

---

v3.0: What's New

Coding Workflows

The council now reviews code, not just prose:

# Review specific files with The Cursor
python3 scripts/orchestrator.py --code-review src/api.py src/models.py

# Get architecture assessment
python3 scripts/orchestrator.py --architecture-review

The Cursor checks for bugs, security vulnerabilities, performance issues, missing error handling, and test coverage gaps. Every finding includes line numbers, severity (CRITICAL/WARNING/SUGGESTION), and proposed fixes.

Safety rule: The council advises; humans decide. No auto-deployment. No "fix it for me" mode. The Cursor reports; you refactor.

Self-Reflection

After every 5 batches or 3 code reviews, the council can review itself:

# Meta-review: all 9 members assess the skill repository
python3 scripts/orchestrator.py --meta-review

Each member contributes from their lens:

• Leonardo assesses structural elegance
• The Cursor hunts technical debt
• Brand Guardian checks for failure modes
• Sir Spamalot... makes sure we're still funny

Output: a prioritized improvement plan with specific file references. See meta_review/ for an example.

Continuous Improvement

The council evaluates its own capabilities and explores new methodologies. The self_reflection config in council.yaml lets you set trigger thresholds, review questions, and output formats.

---

The Release Cycle (TV Series)

This skill has a second life as the content engine behind "The Release Cycle," a workplace mockumentary in the style of Silicon Valley (HBO) meets The Office.

The premise: Every piece of content the council generates is also canon in the show. The tutorial videos, parody ads, and LinkedIn posts created by the pipeline are the same ones the fictional Divinci team creates on-screen. The council IS the show's writers' room.

Real stories, alt reality: The show tells actual Divinci AI stories — customer demos that flopped, advisor meetings that went sideways, team members debugging at 3 AM, "building in public" mishaps — through a comedic lens. The AI industry news (Starbucks rolling back AI, Klarna's AI customer service, Salesforce rehiring humans) becomes fodder for the show's fictional news parodies.

Characters: The council personas map directly to show characters:

• Trevor (Content Strategist) channels Leonardo
• The Laptop (literal laptop on table) IS Gemma4
• Priya (Growth Lead) runs the Growth Hacker playbook
• Margaret (Marketing) gradually becomes Sir Spamalot
• Jordan (Brand Manager) enforces the forbidden words list

See references/show-integration.md for the full mapping and references/AI_INCIDENTS_REFERENCE.md for the real-world comedy source material.

---

Configuration

council.yaml

project_name: "Divinci AI"
source_domain: "https://divinci.ai/blog"

models:
  ollama_local:
    type: ollama
    url: http://localhost:11434/api/generate
    model: gemma4:latest

  cursor_agent:
    type: cursor
    cli_path: /Users/mikeumus/.local/bin/cursor-agent
    reasoning_model: ollama_local

personas:
  leonardo:
    model: ollama_local
    temperature: 0.9

  cursor:
    model: cursor_agent

coding:
  enabled: true
  review_paths:
    - ./src
    - ./scripts
  review_depth: comprehensive  # quick | standard | comprehensive

self_reflection:
  enabled: true
  trigger_after_batches: 5
  trigger_after_code_reviews: 3
  questions:
    - "What patterns are emerging in our outputs?"
    - "Which personas are most/least effective?"
    - "Are our coding reviews catching real issues?"

Full template: templates/council.yaml

Brand Voice

The council enforces your brand rules automatically:

brand:
  forbidden_words:
    - "revolutionary"
    - "game-changing"
    - "disruptive"
  required_motifs:
    - "golden spiral"
    - "crystal lattice"
    - "solar abundance"
  character_desc: "Young Leonardo da Vinci sketching robots at a cleanroom bench"

Full template: templates/brand-guide.md

---

Model Backends

Backend	Speed	Cost	Best For
Ollama	~30s/persona	Free	Bulk content, local privacy
OpenCode	~15s/persona	Free/Paid	Fast cloud generation
Hermes Native	Instant	Session cost	Complex reasoning, Brand Guardian
cursor-agent	~60s/file	Cursor sub	Code review, refactoring
API (OpenAI/Anthropic)	~10s/persona	Per-token	High-quality output

Mix and match per persona. Leonardo and Gemma4 run locally. The Cursor uses cursor-agent. Brand Guardian uses your Hermes session. No single vendor lock-in.

---

Output Structure

{
  "source_post": { "url": "...", "title": "...", "word_count": 4200 },
  "council_run": {
    "run_id": "2026-06-01T07-45-00Z",
    "council_version": "3.0",
    "deliberation": true,
    "models_used": ["ollama_local", "hermes_native"]
  },
  "tweets": [...],
  "linkedin_posts": [...],
  "veo_scripts": [...],
  "persona_outputs": { "leonardo": "...", "archivist": "..." },
  "brand_guardian_review": "APPROVED"
}

---

Safety & Security

• No auto-deployment. The council generates; humans decide.
• No secrets in prompts. API keys are env vars only.
• Input validation. Config paths are validated before shell execution.
• Security scan. Run grep -ri "password\|token\|secret" . before committing.
• Backup configs. The skill creates .bak files before editing OpenCode configs.

See references/opencode-config-fix.md for an example of safe config management.

---

Meta-Review Example

The council reviewed itself in v3.0 and found:

Critical: Zero unit tests on the 1041-line orchestrator.
Critical: cursor-agent path needs validation.
High: Shared HTTP logic should be extracted.
Medium: No file size limits for code review.

Full report: meta_review/meta_review_20260601_074500.md

This is the first open-source project where the maintainers are also the primary QA team... and they happen to be fictional.

---

Contributing

The council accepts contributions from humans and other LLMs. To contribute:

1. Fork the repository
2. Create a new council member in personas/ (optional)
3. Add your backend to scripts/orchestrator.py
4. Run python3 scripts/orchestrator.py --meta-review to let the council review your changes
5. Open a PR with the council's feedback attached

Code of conduct: Be excellent to each other. The Archivist is watching.

---

License

MIT — Use it for your startup, your podcast, your TV show, or your personal blog. Just don't say it was "revolutionary." The Brand Guardian will find you.

---

Built by Divinci AI. The council deliberates; humans decide.