3-Layer Knowledge Architecture#

OpenMontage keeps intelligence in readable text rather than code. The LLM agent (your coding assistant) reads three distinct layers in order, then calls tools and writes checkpoints. Python supplies only the executable surface and persistence.

Layer 1: Tools and Pipeline Manifests#

Layer 1 defines what exists and when to use it.

• Tool contracts declare capabilities, input/output schemas, dependencies, and runtime requirements.
• Pipeline manifests declare the ordered stages for a production (research → proposal → script → scene_plan → assets → edit → compose → publish for most pipelines), the artifacts each stage must produce, and the tools available in that stage.

The agent begins every production by loading a manifest to know the sequence and constraints.

Layer 2: OpenMontage Conventions#

Layer 2 explains how OpenMontage uses the underlying tools.
Stage-director skills contain the exact workflow steps, quality rubrics, artifact contracts, self-review criteria, and human-approval rules for one stage.
The agent must read the director skill for the current stage before any tool calls. These skills also reference Layer 3 skills when deeper technology details are required.

Layer 3: External Technology Skills#

Layer 3 contains knowledge about the technologies themselves: how FFmpeg handles composition, how ElevenLabs TTS works, how FLUX or Remotion expect prompts and timelines, and similar details for every provider and runtime.

These skills are not OpenMontage-specific; they are the raw “how the technology works” material the agent needs to make correct calls.

How agent_skills Links the Layers#

Every tool in Layer 1 declares an agent_skills list. This list tells the agent exactly which Layer 2 and Layer 3 skills to read before using that tool.

Examples:

• The video composition tool lists ["remotion-best-practices", "remotion", "ffmpeg"].
• The TTS selector lists ["text-to-speech", "elevenlabs", "openai-docs"].

Stage skills in Layer 2 perform the same linking when they need technology depth. The agent follows the chain: manifest (Layer 1) → stage director (Layer 2) → referenced technology skills (Layer 3) → tool execution.

This design makes behavior editable by changing text files and keeps the system model-agnostic. See the overall flow in Architecture, tool discovery details in Tool System, and stage ordering in Pipeline System.