6.4 KiB
Loki Mode Voice-Over Script
Complete narration for Loki Mode demo video.
Introduction (0:00 - 0:30)
Welcome to Loki Mode - a multi-agent autonomous startup system for Claude Code.
Loki Mode takes your product requirements document and transforms it into a fully functioning application - with zero human intervention.
Today I'll show you how it works by building a complete todo application from scratch.
Setup (0:30 - 1:00)
First, we launch Claude Code with the dangerously-skip-permissions flag. This allows Loki Mode to run autonomously without asking for confirmation at every step.
[Show terminal:
claude --dangerously-skip-permissions]Now we invoke Loki Mode with our PRD.
Invocation (1:00 - 1:30)
[Type: "Loki Mode with PRD at examples/simple-todo-app.md"]
Loki Mode immediately begins the RARV cycle - Reason, Act, Reflect, Verify.
It first reads the PRD to understand what we're building.
Bootstrap Phase (1:30 - 2:30)
Notice Loki Mode is now in the Bootstrap phase. It's setting up the project structure.
[Show: .loki directory being created]
The .loki directory contains:
- CONTINUITY.md - the working memory that persists across context resets
- Queue files for task management
- State tracking for the orchestrator
This is how Loki Mode maintains context even during long-running operations.
Discovery Phase (2:30 - 3:30)
Now we're in Discovery. Loki Mode is analyzing our PRD and extracting requirements.
[Show: Tasks being generated]
See how it breaks down the todo app into specific tasks:
- Set up backend with Express
- Create SQLite database schema
- Implement API endpoints
- Build React frontend
Each task gets added to the pending queue.
Architecture Phase (3:30 - 4:30)
The Architecture phase is where Loki Mode designs the system.
[Show: OpenAPI spec being created]
Notice it's following spec-first development - the OpenAPI specification is created BEFORE any code is written.
This ensures the frontend and backend will work together seamlessly.
Kanban Visualization (4:30 - 5:30)
Let me show you the Vibe Kanban integration.
[Show: Kanban board with tasks]
Each task appears on our kanban board. As agents claim tasks, they move from "To Do" to "In Progress" to "Done".
This gives you real-time visibility into what Loki Mode is doing.
Agent Spawning (5:30 - 7:00)
Now watch the magic happen.
[Show: Multiple agents being spawned]
Loki Mode spawns specialized agents:
- A backend agent implementing the Express server
- A frontend agent building the React UI
- A database agent setting up SQLite
These agents work in parallel - but notice they're not stepping on each other's toes. The task queue system prevents conflicts.
Model Selection (7:00 - 7:30)
Pay attention to the model selection.
Simple tasks like running tests use Haiku - fast and cost-effective. Standard implementation uses Sonnet - the default workhorse. Complex decisions like architecture use Opus - for deep analysis.
This intelligent routing optimizes both speed and quality.
Code Review (7:30 - 9:00)
Here's my favorite part - the code review system.
[Show: Three reviewers being dispatched]
Loki Mode dispatches THREE reviewers in parallel:
- Code quality reviewer - checks patterns and best practices
- Business logic reviewer - verifies requirements are met
- Security reviewer - scans for vulnerabilities
They review independently - blind to each other's findings. This prevents groupthink.
[Show: Review results]
If all three approve, a Devil's Advocate reviewer is triggered. This fourth reviewer specifically looks for issues the others might have missed.
This anti-sycophancy protocol catches 30% more issues than traditional reviews.
Quality Gates (9:00 - 10:00)
Severity-based blocking ensures nothing ships broken.
[Show: Quality gate output]
Critical, High, and Medium issues BLOCK the pipeline. Low and Cosmetic issues get TODO comments but don't block.
Tests must pass. Coverage must exceed 80%. No exceptions.
CONTINUITY.md (10:00 - 11:00)
Let's peek at the working memory.
[Show: CONTINUITY.md contents]
This file tracks:
- Current task and progress
- Decisions made and why
- Mistakes and learnings
If Loki Mode runs out of context or needs to restart, it reads this file first. This is how it maintains coherence across long sessions.
Memory System (11:00 - 12:00)
Loki Mode has a three-layer memory system.
Episodic memory records what happened - specific actions and their outcomes.
Semantic memory generalizes patterns - "TypeScript strict mode requires explicit return types."
Procedural memory stores learned skills - how to implement an API endpoint successfully.
This isn't just context - it's genuine learning that improves future runs.
Completion (12:00 - 13:00)
[Show: Application running]
And here's our finished todo app!
- Full CRUD operations working
- React frontend with TypeScript
- Express backend with SQLite
- All tests passing
- Code reviewed and approved
From PRD to working application - completely autonomous.
Recap (13:00 - 14:00)
Let's recap what Loki Mode did:
- Read and analyzed the PRD
- Designed the architecture with OpenAPI specs
- Spawned specialized agents for parallel development
- Ran comprehensive code reviews with anti-sycophancy checks
- Enforced quality gates and test coverage
- Maintained context through the memory system
All without a single human intervention.
Call to Action (14:00 - 14:30)
Loki Mode is available now on GitHub.
Install it as a Claude Code skill and start building.
Remember to use the dangerously-skip-permissions flag for full autonomy.
Thanks for watching!
Timing Summary
| Section | Start | Duration |
|---|---|---|
| Introduction | 0:00 | 30s |
| Setup | 0:30 | 30s |
| Invocation | 1:00 | 30s |
| Bootstrap | 1:30 | 60s |
| Discovery | 2:30 | 60s |
| Architecture | 3:30 | 60s |
| Kanban | 4:30 | 60s |
| Agents | 5:30 | 90s |
| Model Selection | 7:00 | 30s |
| Code Review | 7:30 | 90s |
| Quality Gates | 9:00 | 60s |
| CONTINUITY | 10:00 | 60s |
| Memory | 11:00 | 60s |
| Completion | 12:00 | 60s |
| Recap | 13:00 | 60s |
| CTA | 14:00 | 30s |
Total: ~14.5 minutes