escriptorium

Read the page's lines and transcription records first, then use create_line_transcription with a segmented line ID and layer ID, or update_line_transcription with the line transcription record ID. The record ID differs from the line ID, so always resolve it from prior reads.

Submit jobs with explicit page IDs and a model whose job type matches: 1 for segmentation, 2 for recognition. Training needs a ground-truth layer plus a starting model or new model name; a successful submission only means queued, so inspect task reports before reporting completion.

No. The skill only directs use of the connected escriptorium MCP tools and does not install or authenticate the server. If the connection is unavailable, it reports that prerequisite instead of attempting transcription work.

Use export_transcriptions for direct text or JSON of a single layer, and request_server_export for native ALTO/PAGE XML, text, or JSON document archives with optional images, annotations, and metadata. OpenITI and TEI formats require server-side enablement.

Bulk line and text operations are non-atomic and can partially apply on failure. After any failure, re-read the affected records before retrying, and pause concurrent writers since preflight membership checks are not locks.

No. The skill is scoped to eScriptorium workflows through its MCP server and is not intended for general handwriting transcription without an eScriptorium connection. For standalone OCR of images, a different tool would be required.