Principle 1: Always Verify First

The Automated Guardrails

"Don't automate the creation before you automate the verification."

• Never allow an agent to write feature code without an automated check.
• Write tests, specs, or canary scripts first.
• Let the agent fail fast against strict local constraints.

Principle 2: Local-First & File-Based

One Source of Truth

• Specs, decisions, issues, and rules live as files in the repo.
• Files like AGENTS.md, CONTEXT.md, and issues/ are accessible to both humans and agents.
• No proprietary cloud dashboards. If it's not in Git, it doesn't exist.

Principle 3: Human in Control

Model Agnostic, Transparent Design

• Frontier models get better every month.
• Heavy agentic frameworks, multi-layer prompts, and resident databases age poorly.
• Keep helper tooling minimal. The human triggers the skills and commands manually.

You can delegate execution — you cannot delegate accountability.

Principle 4: Canary-First Verifications

Probe the Environment Before Trusting It

• External CLI tool? OS-level behavior? Sandbox policy? Build a canary first.
• A thirty-line script in the real environment beats a weekend of debugging.
• Run it, observe, document the finding — and keep it as a regression check.

Principle 5: Keep It Minimal and Radical

Every Line You Don't Write

• Simple, clean, readable code over generic, bloated patterns.
• Self-contained static binaries over resident daemon architectures.
• House Rules only — a subset chosen so a reviewer, or a check, can tell right from wrong fast.
• One component, one language. A second implementation is a divergence waiting to happen.

A line you don't write is a line no agent can misunderstand.

The Agentic Dev Cycle

A Spec-Driven, Always-Verify Loop

graph TD
    Spec[1. Write/Update Spec/Glossary] --> Harness[2. Create Verification Harness/Canary]
    Harness --> Prompt[3. Execute Agent Command]
    Prompt --> Verify{4. Checks Pass?}
    Verify -- No --> Prompt
    Verify -- Yes --> Git[5. Commit & Push]

Summary

The Ubunatic Way

1. Verify first: Control the output through automated checks.

2. File-based: Write decisions down where the agent can read them.

3. Stay in control: The human triggers; the agent executes.

4. Canary-first: Probe every external mechanism before building on it.

5. Be minimal: Limit complex dependencies; hold to the House Rules subset.