Set up task management. Parallelize with sub-agents. Ensure resumability after interruption.
Build a manifest of all input documents listing file type and size.
Convert non-text files to readable text with minimal content loss. Prefer existing tools (e.g. pandoc) over AI models, which cost more tokens. Use a cheap or local multimodal model for image descriptions. Mark empty or unconvertible files as skipped.
Categorize documents into tiers by content type, size, and complexity. Store tiers in a data file. Ask the user to clarify size thresholds (bytes, characters, tokens, or lines) or expected modules if ambiguous.
- Tier 1: Complex, small-to-medium, or high-priority documents. Use the most capable AI model.
- Tier 2: Medium documents with simple, repetitive, or low-priority content. Use a cheaper, faster model.
- Tier 3: Large or highly repetitive/low-priority documents. Use the cheapest model or programmatic summarization.
Pre-process file types programmatically without per-file AI calls:
- Generate initial source code descriptions with language-appropriate documentation tools.
- Extract structural metadata from each file (e.g. parent class, interfaces, abstract/concrete, fields, members, properties). Define a consistent header format with a fixed item order across all description files.
- Batch-describe images via a cheap or local multimodal model.
- Produce one Markdown description file per non-skipped document.
- Write resumable pre-processing scripts with a formal execution plan.

handbook-writer