Arabic Calligraphy for AI Agents
An open-source toolkit I built for Arabic typography, editable Square Kufic studies and local artwork exports through AI coding agents.
I built Arabic Calligraphy for AI Agents to give coding agents a practical, reviewable way to produce Arabic artwork. It combines a portable Agent Skill, an offline Python renderer and editable geometric studies, with the original Unicode text retained throughout the workflow.
Source repository: github.com/akin-demir/arabic-calligraphy-skill.
From an instruction to an actual file
The toolkit supports four font workflows: Naskh, modern Kufi, Nastaliq and limited Diwani. Agents can create wordmarks, posters, certificates and new documents, then deliver real PNG, outlined SVG or PDF files where the selected workflow supports them. Pango and HarfBuzz handle Arabic shaping and right-to-left layout; Cairo handles export. Rendering runs locally after installation.
The reusable skill explains the supported capabilities and failure cases so an agent can report a missing dependency or unsupported request. Source, installable packages, editable examples and checksummed release assets are available publicly.
Square Kufic as editable geometry
The Square Kufic workflow uses manually authored, labeled grid cells. Each letter maps back to a span in the original text, making the intended reading available for review alongside the artwork. Geometry checks flag uneven blocks and corner-only contacts before export.
The cover is generated from the project’s actual grid artwork: four studies of
الله rotated around a central space, with teal and gold distinguishing the
copies. The Unicode source is preserved; no generic font is pixelated and no
filler strokes are added to imitate Arabic lettering.
Current scope
RC4 is an experimental public release, with 482 automated checks recorded in its release evidence. The four font workflows are approximations, and the manual compositions have not received independent calligrapher review. Arbitrary-text Square Kufic conversion and automatic packed composition remain future work. Diwani currently supports unvocalized PNG/SVG only; Arabic PDF extraction, search and copy behavior still varies by reader.