AI TOOLS · HANDWRITING · AGENT SKILL
Inko Handwriting Skill: Turning Agent Output into Editable Pen Strokes
From character validation and authentication to flat PNG/PDF delivery: an open-source workflow for controlled handwriting generation
Source: pretend1111/inko-skill · pretend1111
I
User Problem
A verifiable delivery chain, not a font filterGeneral image models can imitate handwriting in a small sample, but they struggle to preserve long text, formulas, pagination, paper specifications, and character-level accuracy. Inko reframes the task as an agent workflow: prepare and validate the content, choose paper and pen settings, generate standard flat pages, then inspect and adjust them locally. The repository’s promise is natural handwriting for notes, letters, essays, and math solutions—not text pasted into a simulated lifestyle photo.Source
- Math solutions: read the problem, validate formulas and symbols, then produce a flat ruled-paper answer sheet.
- Notes: lay out notes.md on grid or white paper while retaining an editable stroke package.
- Letters and essays: choose cream paper, composition paper, or practice grids and deliver PNG plus PDF.
* * *
II
Product Mechanism
Generation · layout · local editing · live capability data- Natural strokes, flat deliveryThe system outputs standard PNG and PDF pages on white, cream, grid, ruled, composition, or character-practice paper. A problem photo may provide input content, but the result remains a standalone flat page. Source
- Layout remains editableTitles, text boxes, line spacing, ink, pen type, weight, and character spacing can be adjusted; scene.py preserves a stroke package whose glyphs can be moved and re-rendered. Each user also receives one free custom-handwriting seat with two training runs.Source
- Local post-processingPaper rendering, ink styling, PDF assembly, and line-position adjustments run through repository scripts locally. The README states that Pillow/numpy image processing does not upload images.Source
- Capabilities are queried liveWritable characters and Logic-model LaTeX commands are no longer hard-coded in the skill. charset and models --symbols query the server so documentation does not silently lag model support.Source
* * *
III
Workflow
Four steps from installation to deliverable pages- Create a dedicated API keyCreate a key specifically for the agent and set a daily spending limit. Handwriting inference, the website’s DeepSeek assistant, and the external agent model are accounted for separately.
- Install at the right scopeClone the repository and run install.py with Python 3.9+. The installer detects Claude Code or .agents-compatible tools and supports explicit project- or user-level scope.Source
- Store the key through stdinUse inko.py auth to read the key from standard input, avoiding command arguments, project files, and commits. It defaults to the user config directory or an ignored project directory.Source
- Check, preview, generate, inspectRun doctor, then use quote or layout --preview to validate characters, formulas, and fit. Submit with --yes only after a zero quote, then open every PNG to inspect missing text, collisions, alignment, and labels.Source
* * *
IV
Limits and Boundaries
Flat output · disclosure · anti-forgery| Boundary | Current repository rule | How to operate |
|---|---|---|
| Flat two-dimensional pages only | Since version 1.3.0, photo realism, desk scenes, perspective, page curl, scan/copy effects, and photo compositing have been removed. The repository description still mentioning phone-photo looks is stale metadata.Source | Treat images only as content input; keep output as independent, flat PNG/PDF pages. |
| Models and characters have limits | Lyric targets prose while Logic handles formulas. Supported characters and symbols change, so old documentation and model memory are not authoritative. | Run charset, models --symbols, and quote before generation; confirm a writable substitute when needed. |
| AI labels must remain | The repository requires a visible “AI生成 · Inko” label and hidden metadata on every page. Only accounts with the relevant agreement may request output without the visible label.Source | Do not crop, cover, or re-save away the metadata; disclose AI generation when sharing. |
| Forgery-like documents are refused | IOUs, receipts, contracts, certificates, leave notes, signatures, and similar documents are blocked; renaming or splitting content is not a valid workaround.Source | Use it for legitimate notes, coursework, letters, and essays; never imitate a real person’s handwriting without consent. |
Scroll horizontally to view the full table
The skill’s real value is not another image-generation button. It links proofreading, character support, price blocking, secret storage, layout preview, visual QA, and compliance into an auditable delivery chain. It fits handwriting tasks that require exact text and continuous pages; it is not a photo-realism engine, signature tool, or review-free document generator.
* * *
Editor's Note
How to Read This Article
Why it matters
Unlike repositories that show only a few attractive handwriting samples, this one specifies installation, authentication, live character support, zero-price blocking, local post-processing, editable strokes, and final inspection as an operating contract. Its decision to make flat pages a hard product boundary is more useful than continuing to chase deceptive photo realism.
Caveats
All capability, pricing, compliance, and privacy claims come from the project’s own README and skill; this review did not independently test API availability, handwriting quality, queue capacity, or legal interpretation. The stale phone-photo wording in the repository description is itself a warning to prefer versioned documentation while treating server-side policy as changeable.
Transferable insight
The same design transfers to any external generation service: query capabilities live, turn price into a machine-readable pre-submit gate, keep secrets in local configuration, and require visual inspection after generation. A strong agent skill converts an unpredictable model into a recoverable, traceable workflow.
Source excerpt