handwrite

Turn your real handwriting into a usable Unicode font (.otf / .ttf / .woff2) you can use on the web, in editors, and in messages — processed on-device, so your handwriting never leaves your machine.

Status: font engine proven; emotional fidelity unproven. This repo contains the portable core library + five runnable proofs. After an adversarial review (see docs/report.html), the honest position is: the font container, variation, curve quality, and spacing are de-risked in code, but whether a real person's handwriting renders as recognisably theirs is the open kill-gate (PLAN.md Phase 0.5). See docs/PLAN.md and docs/report.html.

The idea

1. Ingest your handwriting — either by drawing with an Apple Pencil / pointer, or by photographing a printed calibration sheet you've filled in (one character per box).
2. Vectorize each glyph into the Bézier outlines fonts are made of.
3. Assemble a real font file, with natural per-letter variation, that works everywhere.

Proofs (run them)

bun install
python3 -m venv .venv && .venv/bin/pip install fonttools brotli uharfbuzz

bun run proof:all     # runs A, C, B, E, D in sequence
bun run validate      # independent fontTools validation of every generated font

# or individually:
bun run proof:font        # A: opentype.js builds a valid .otf in pure JS  (browser path)
bun run proof:stroke      # C: pen stroke -> perfect-freehand -> font glyph (direct-draw path)
bun run proof:variation   # B: fontTools real .ttf + .woff2 + calt cycling  (variation path)
bun run proof:curvefit    # E: RDP+cubic-fit (127→59 cmds) + ink-derived metrics
bun run template          # D: generate a printable calibration sheet (photo path front-end)

Proof	Proves	Result
A 01-opentype-font.ts	JS-only, in-browser → valid OpenType font (lines, curves, counters, cmap)	✅ valid CFF .otf
C 02-stroke-to-glyph.ts	captured pen stroke → smooth glyph outline → font	✅ stroke→glyph
B 03-calt-variation.py	real .ttf + .woff2 + calt alternates that shape/cycle in HarfBuzz	✅ a → a.calt1 → a.calt2 → …
D 04-generate-template.ts	printable calibration sheet with registration markers	✅ SVG sheet
E 05-curvefit-metrics.ts	(post-review) RDP+cubic-fit kills point-bloat; metrics from ink	✅ 127→59 cmds; advances vary

Repo layout

src/lib/handwrite/        portable, framework-agnostic core (the future app imports this)
  types.ts                shared GlyphOutline/FontSpec data model (both ingest modes target it)
  font/opentype-builder.ts  FontSpec -> opentype.js Font -> bytes (browser)
  capture/stroke.ts       pointer strokes -> perfect-freehand outline -> normalized glyph
proofs/                   runnable feasibility proofs + fontTools validator
docs/                     PLAN.md, ARCHITECTURE.md, RESEARCH.md, report.html

Key decisions (see docs/ARCHITECTURE.md)

• Browser-first, on-device — privacy moat + ~$0 infra (Cloudflare Pages).
• opentype.js writes the font in the browser (a valid CFF .otf); a fontTools step (also runnable client-side via Pyodide) adds true .ttf/.woff2 and the calt variation feature opentype.js cannot author.
• Variation = multiple alternates + deterministic calt cycling, not variable fonts, not rand.
• Photo path requires a printed grid + fiducial markers — that's what makes segmentation tractable.
• GPL note: Potrace and its WASM/JS ports are GPL-2.0 — resolve before monetizing (service boundary, commercial licence, or a permissive tracer).