document-enhancer

name	document-enhancer
description	Enhances and hardens markdown documents through two independent pipelines — (1) fusing external knowledge bases into a target document via multi-agent orchestration, and (2) recursively optimizing documents for multi-model LLM effectiveness using a split→optimize→review→merge architecture.

Document Enhancer

A dual-pipeline system for improving markdown documents. Both pipelines use the same underlying multi-agent architecture (Orchestrator → Sub-Agent → Reviewer → Merge), but differ in their purpose and inputs.

Pipeline 1: Knowledge Fusion (fuse_knowledge.sh)

Fuses external knowledge (e.g., a video knowledge base produced by the video-knowledge-extractor skill) into a target document.

When to Use

• You ran the video-knowledge-extractor and have a _analysis.md knowledge base ready
• You have any external reference material you want woven into a document
• You want to enrich a document with factual, sourced intelligence

How it Works

1. Condense: An Orchestrator agent reads and condenses the knowledge base, extracting only high-signal findings relevant to the target document.
2. Split: The target document is split into semantic sections (on H2/H3 headers).
3. Plan: The Orchestrator identifies which sections should receive knowledge and produces an integration plan (JSON).
4. Integrate: A Sub-Agent weaves the relevant knowledge into each targeted section — surgically, preserving existing content.
5. Review: A Reviewer agent validates each enhanced section against a Quality Assurance Gate (fidelity, utility, token discipline).
6. Reassemble: Approved sections are merged back into the final document, with a diff produced.

Usage

.agents/skills/document-enhancer/fuse_knowledge.sh "path/to/target_document.md" "path/to/omni_output_dir"

• First argument: target document to enhance
• Second argument: directory containing _analysis.md knowledge base files (output from video-knowledge-extractor)

QA Validation Gate

The Reviewer evaluates each enhanced section on:

• Fidelity: Are original objectives and domain terms preserved?
• Effectiveness: Is the integrated knowledge woven naturally, not just appended?
• Token Discipline: Is section growth surgical (under 30%)?

---

Pipeline 2: Recursive Document Hardening (harden_document.sh)

Recursively optimizes any markdown document for maximum effectiveness across all major LLM families (GPT, Claude, Gemini, Llama, Mistral).

When to Use

• You have a prompt, system instruction, or agent definition that needs hardening
• You want to improve token efficiency without losing semantic density
• You need a document to parse reliably across different model architectures

How it Works

1. Split: Document is split into semantic sections on H2/H3 headers.
2. Optimize: Each section is dispatched to a specialized Sub-Agent that applies multi-model formatting (XML tags for Claude, delimiters for GPT, hierarchical markers for Gemini, clear markdown for open-source models).
3. Review: A Reviewer agent scores each optimized section. Sections below threshold are flagged for revision.
4. Iterate: The pipeline loops up to 5 times — sections that pass are locked; sections that fail are re-optimized with reviewer feedback injected.
5. Merge: Final optimized document is assembled and a diff is generated.

Usage

.agents/skills/document-enhancer/harden_document.sh "path/to/target_document.md"

Control iterations via environment variable:

MAX_ITERATIONS=3 .agents/skills/document-enhancer/harden_document.sh "path/to/doc.md"

---

Architecture Constraints (Both Pipelines)

• Context Pollution Prevention: The full document is never sent to a single agent. Both pipelines handle semantic chunking natively. All agents are injected with the project's AGENTS.md structured reasoning framework to prevent context drift.
• Camofox Web Stealth Integration: The sub-agent and orchestrator system prompts include the camofox-stack SKILL.md, giving agents native expertise to write and execute stealth CLI web-scraping scripts if external intelligence is needed.
• Provider Fallback: All API calls route through lib/provider.sh, which automatically falls back between OpenRouter, ZenMux, and Nebius if a provider drops a request.
• Do not edit the scripts manually to fix context limits; the pipelines already handle chunking, dispatch, and reassembly.

Environment & Configuration

A .env file must be configured. Copy .env.example from the skill directory.

If the pipeline fails:

1. Verify .env exists.
2. ORCHESTRATOR_PROVIDER, SUBAGENT_PROVIDER, REVIEWER_PROVIDER must be configured.
3. For the knowledge fusion pipeline, OMNI_PROVIDER and OMNI_MODEL are also needed (set by video-knowledge-extractor).