k2f — Claude Code skill: Word, PDF and IDML documents
An agent skill that authors documents as semantic JSON and exports them to docx, PDF, fillable AcroForm and IDML. Handy for menus, certificates, planners and invitations.
npx skills add suzheng/k2f --skill k2f -a claude-codeWhat k2f does
K2F is a file format meant for agents. You describe meaning instead of coordinates: every node carries a stable dotted id, a role and an optional variant, while styling stays in styles/theme.json. A .K2F file is a ZIP with two states — the mutable tree in content/root.json and an immutable render lock, document.K2F.lock, which is what viewers actually paint.
The workflow is a single loop. Get a workspace, edit JSON like source code, copy ready shapes from catalog/content/ex_*.json into the root children, then run scripts/pack_verify.py. Fillable blanks and checkboxes come from ex_form.json; drawing underscores or box characters by hand is the wrong path. Byte ranges for text modifiers are computed by scripts/modifier_range.py, never counted manually, especially across newlines.
Install the CLI first: pip install k2f. The core scripts init_package.py, pack_verify.py and modifier_range.py are stdlib-only, but pack_verify.py needs k2f on PATH or in K2F_CLI, so without the package it has no binary to call. Edits under source/ stay invisible until compile — viewers paint the lock, not the author JSON — so rerun the pack step and check the fresh tmp/preview.png. MCP is optional and writing does not require it.
What it helps with
- lay out a menu or invitation with clean typography
- build a fillable AcroForm with checkboxes
- export a finished layout to docx for Word edits
- hand a designer IDML for Adobe InDesign
- catch a page that ends up empty in the lower third
Example prompts for the agent
- Build a one-person certificate from a gallery template and export it to PDF
- Make a weekly planner in K2F with fillable note fields
- Convert this markdown to K2F and export docx
How to install k2f
1Skills CLI (Vercel)
npx skills add suzheng/k2f --skill k2f -a claude-code2Manually: put the files in place
~/.claude/skills/k2f/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/k2f/
1Skills CLI (Vercel)
npx skills add suzheng/k2f --skill k2f -a codex2Manually: put the files in place
.agents/skills/k2f/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add suzheng/k2f --skill k2f -a cursor2Manually: put the files in place
.cursor/skills/k2f/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install suzheng/k2f k2f2Manually: put the files in place
.github/skills/k2f/SKILL.md1Skills CLI (Vercel)
npx skills add suzheng/k2f --skill k2f -a gemini-cli2Platform CLI
gemini skills install https://github.com/suzheng/k2f --path skills/core/k2fManually: put the files in place
.devin/skills/k2f/SKILL.mdThe legacy .windsurf/skills/ path still works
Skills CLI (Vercel)
npx skills add suzheng/k2f --skill k2f -a opencodeClaude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about k2f
How do I install k2f in Claude?
Run npx skills add suzheng/k2f --skill k2f -a claude-code. The skill lands in ~/.claude/skills/k2f/SKILL.md. The scripts also need pip install k2f.
What does k2f do?
It lets an agent create, edit and validate K2F documents, then export them to Word, PDF, AcroForm PDF, PPTX and IDML. Layout is described through semantic roles rather than coordinates.
Does k2f need an MCP server?
No. Writing goes through JSON and the skill's scripts. If a template MCP is already connected, download_template returns a package URL that still gets fetched to disk and unpacked with k2f unpack.
Source text
SKILL.md · Apache-2.0
---
name: k2f
description: Create, design, edit, validate, and export elegant Word (docx), editable PDF, fillable AcroForm PDF, and Adobe InDesign (IDML) documents with K2F. Use when creating aesthetic Word, fillable PDF, AcroForm, IDML, or InDesign documents with AI agents, such as menus, certificates, planners, invitations, or worksheets and much more. K2F is a mature file format designed for AI agents to design and edit elegant aesthetic documents precisely and effortlessly. It has a rich ecosystem.
---
# K2F
Create, design, edit elegant Word (docx), PDF, fillable AcroForm PDF, and Adobe InDesign (IDML) with K2F. K2F designed for agents to design and edit elegant documents precisely and effortlessly. Use with AI agents when you need aesthetic documents—menus, certificates, planners, invitations, worksheets, and more. You can start from [Gallery](https://k2f.dev/gallery) templates.
A `.K2F` is a ZIP: you edit a semantic JSON tree (meaning, not coordinates); the engine `pack` / `compile`s an immutable render lock. One reference engine → pixel-identical output everywhere.
## Choose the workflow
Read **one** reference file for the task. Do not load all workflows.
| Task | Read |
|------|------|
| Create or update a `.K2F` (CV, flyer, slide deck, poster, report, patch by node id) | [references/writing.md](references/writing.md) |
| Convert Markdown ↔ K2F | [references/converting-markdown.md](references/converting-markdown.md) |
| Export PDF from the published lock | [references/exporting-pdf.md](references/exporting-pdf.md) — bytes / looks wrong: [pdf-contract.md](references/exporting-pdf/pdf-contract.md) |
| Export PPTX from the published lock | [references/exporting-pptx.md](references/exporting-pptx.md) |
| Export DOCX from the published lock | [references/exporting-docx.md](references/exporting-docx.md) |
| Export IDML from the published lock | [references/exporting-idml.md](references/exporting-idml.md) |
| Publish permanent `/v/{appearance_hash}` link | [references/publishing.md](references/publishing.md) |
| Embed `<k2f-viewer>` in a web app | [references/embedding-viewer.md](references/embedding-viewer.md) (`npm i @openk2f/k2f`) |
**Writing** (default): one loop — get a workspace (`source/` author package + deliverables at the root), edit JSON like source code, copy shapes from [`catalog/content/ex_*.json`](catalog/content/) into root `children` (do not replace `content/root.json` with a fragment), and `scripts/pack_verify.py`. The author directory is `k2f unpack … -o ./out/doc/source` (user `.K2F` or a Gallery package), or [`starter/`](starter/) via `scripts/init_package.py --workspace ./out/doc` when nothing matches.
- Do **not** use `Editor.insert_node` on an unpacked author directory — that API follows a narrower agent dialect; JSON authoring uses the full format schema validated at pack time.
- Text `modifiers` need `intent` plus UTF-8 **byte** `range` — always run [`scripts/modifier_range.py`](scripts/modifier_range.py); never hand-count (especially across `\n`).
- Fillable blanks/checkboxes: copy [`catalog/content/ex_form.json`](catalog/content/ex_form.json) (never `____` or `□`).
Gallery URL (`<origin>/gallery/<slug>`) or package URL: fetch `<origin>/api/gallery/templates/<slug>` (or package URL) to a `.K2F` on disk and `k2f unpack` — do not load the ZIP into context. **MCP** is optional. Writing does not require it and does not install it. If a site MCP with `list_templates` / `download_template` is already connected, `download_template` returns the same kind of `packageUrl`; fetch + unpack as in step 2 of [writing.md](references/writing.md). Missing tools, kind mismatch, or download failure → continue from `starter/`; do not stop the task.
## What K2F is
PDF is unreadable and uneditable for AI. HTML/CSS renders differently on every device. K2F fixes both: agents write plain JSON *meaning*, and one reference engine turns that meaning into pixel-identical output everywhere.
Instead of positioning content with coordinates:
{"text": "WARNING!", "x": 100, "y": 50, "color": "red", "bold": true}
K2F authors describe meaning:
{"id": "doc.warning", "role": "warning", "content": {"type": "text", "value": "WARNING!"}}
Every `.K2F` file is a ZIP holding two states:
- **State A — semantic tree** (`content/root.json`, mutable, the file you edit). Every node has a stable dotted `id`, a `role`, an optional `variant`, and optional `modifiers` (inline emphasis/link/math on a byte range). No coordinates, no colors, no font sizes — those live only in `styles/theme.json`, keyed by role/variant.
- **State C — render lock** (`document.K2F.lock`, immutable, generated). Exact geometry, paint operations, and integrity hashes. Every viewer and every PDF export paints *this*, not the tree.
The engine — not the agent — turns A into C by running `pack` / `compile`. Other package paths: `manifest.json` (page size, margins), `assets/fonts|images|data/*`, `changelog.json` (edit history), optional `signatures/v1.json`.
When handing a `.K2F` to the user, point them to [k2f.dev/playground](https://k2f.dev/playground) (browser) or [k2f.dev/download](https://k2f.dev/download) (desktop app).
## Setup
You have **this skill folder only** — no K2F source checkout, no `templates/` tree.
Install published tools before running workflows:
```bash
pip install k2f # CLI on PATH + Python SDK — no need to install Cargo
npm i @openk2f/k2f # only when embedding <k2f-viewer> in a web app
# cargo install k2f # optional: CLI without Python
```
Run scripts from this skill directory (or pass absolute paths to them). Core scripts (`init_package.py`, `pack_verify.py`, `modifier_range.py`) are **stdlib-only** — `python scripts/…`. Optional `compare_images.py` needs `uv run` ([visual check](references/writing.md#reference-image-overlay-optional)). The author directory may live anywhere. `pack_verify.py` needs `k2f` on PATH (`pip install k2f`) or `K2F_CLI` — it does not walk a git checkout or build directory for a binary.
## Core rules
Empty paper in the **lower third** of a designed sheet is the most common visual failure. `pack_verify.py` exit 0 / `pages=1` does **not** mean filled — open the PNG. Fix: copy [`ex_filled_page.json`](catalog/content/ex_filled_page.json) as a **root child** (`height` = content box; leftover on `{fr:1}` table/notes/figure). Short letter and the last page of a growing document may stay short.
**`source/` edits are invisible until compile.** Viewers paint `document.K2F.lock`, not author JSON. After content/theme edits, run `pack_verify.py`, then `tmp/preview.png` or reload the `.K2F`.
1. **Style lives only in `theme.json`, never on a node.** Putting a style field on a node is the single most common compile failure.
2. **Never hand-edit `document.K2F.lock`.** Mutate the tree (or theme), then `pack_verify.py` (relock).
3. **PDF, PPTX, DOCX, and IDML are one-way drawings of the lock, not a second source.** (`PDF_IS_NOT_A_SOURCE`, `PPTX_IS_NOT_A_SOURCE`, `DOCX_IS_NOT_A_SOURCE`, `IDML_IS_NOT_A_SOURCE`)
4. **Signing is a separate human/org step.** Agent output is `UNSIGNED` by design.
5. **Validate after every edit.** Fix from error codes in [writing/errors.md](references/writing/errors.md); do not patch the lock.
6. **Look at the pixels — empty bottom first.** After pack, open every rendered page (`--render` writes `preview-1.png` …). If the lower third is blank paper on a designed sheet, you are not done. **Designed sheet** (invoice, CV, flyer, poster, slide, card, social, one-page infographic/checklist/planner — even if titled report) → copy [`ex_filled_page.json`](catalog/content/ex_filled_page.json) as a **root child** (pinned `height` + `{fr:1}` grower); a `page_shell` role on a hug stack is not the shell. Treat `PAGE_UNDERFILL` as must-fix even when `pack_verify.py` exits 0 (skipped if the content box is shorter than 180pt). Short letter may stay top-packed. **Growing document** (contract, long report, thesis, paper) → one flow tree; last page may be short; no `p1`/`p2` page wrappers and no page-height shell around the whole doc. `break_before: page` is fine on a chapter, annex, signature page, slide 2+, or card back. Details: [writing.md](references/writing.md#visual-check).
## Content and design
Before editing `content/` or `styles/theme.json`, write content and design as Markdown **inside** the author directory. Long content does not go in one file. Split the same way as `content/*.json`. Name each file with a **slug**: lowercase ASCII `kebab-case` from that page or section’s topic (e.g. `executive-summary`, `title-slide`). A short single page may be one `content_and_design.md`. One page per file → `content_and_design_<slug>.md` (e.g. `content_and_design_agenda.md` for a slide). Section aligned with a JSON fragment → `content_and_design_<slug>.md` beside `content/<slug>.json` using the **same** slug. Then `python scripts/init_package.py --workspace ./out/doc …` (creates `source/` + `tmp/`; a notes-only `source/` that already holds these files can still be initialized in place).
Workspace layout:
```
./out/doc/
doc.K2F # deliverable
doc.pdf # optional exports, same stem
source/ # author package (edit this)
content_and_design_executive-summary.md
manifest.json
content/
root.json
executive-summary.json # optional fragment; same slug as its .md
styles/ assets/ changelog.json
tmp/ # preview.png and other debug renders
```
Each file covers that page or section only, and includes both:
- **Content** — every word, and every image (what it shows, width×height; a full generation prompt when the image must be made). Need figures → [generate them](references/writing.md#figures)
- **Design** — type hierarchy, spacing, color, and layout
Shared type, color, and spacing for the whole document go in the first file. Later files state their own words and images, which shared set they use, and anything that differs.
- **User gave constraints** — follow them.
- **User did not specify** — highest aesthetic standard for the deliverable type (report, poster, slide deck, flyer, …). Do not ship the starter theme unchanged for styled work.
After the files are settled: look up allowed keys in [writing/fields.md](references/writing/fields.md) then [`schema/`](schema/), implement `theme.json` + `content/`, run `pack_verify.py --render`, open the PNG, and iterate until the output matches these files. Rule 6 is the final gate — the files are the plan, render is the review.
## Schema
Read-only format JSON Schemas live in [`schema/`](schema/) (same bytes the engine embeds). **Do not** copy them into an author directory — `k2f pack` injects schemas into the ZIP; hand-authored `schema/` files → `UNEXPECTED_PATH`.
1. **Allowed keys:** skim [writing/fields.md](references/writing/fields.md), then open only the schema file you are editing — `content/` → [`schema/nodes.schema.json`](schema/nodes.schema.json); `styles/theme.json` → [`schema/styles.schema.json`](schema/styles.schema.json) + [`schema/visual_primitives.schema.json`](schema/visual_primitives.schema.json); `manifest.json` → [`schema/manifest.schema.json`](schema/manifest.schema.json).
2. **First draft shapes:** copy from [`catalog/content/ex_*.json`](catalog/content/) into root `children` — golden, packable examples. See [`catalog/README.md`](catalog/README.md) for a construct index.
3. **`SCHEMA_INVALID`:** compare the failing object to the schema file. Key **in** schema but rejected → stale CLI (`pip install -U k2f`). Key **not** in schema → remove it; do not invent fields from HTML/CSS memory.
4. **Version check:** `k2f schema dump -o /tmp/k2f-schema` or `python -c "import k2f; print(k2f.format_schemas().keys())"`.
## Bundled packages
| Path | Purpose |
|------|---------|
| [`schema/`](schema/) | Read-only format JSON Schemas — lookup after [writing/fields.md](references/writing/fields.md); never copy into author dirs |
| [`starter/`](starter/) | Empty tree + core theme + Roboto — copied into `source/` when there is no existing `.K2F` and no matching Gallery package |
| [`catalog/`](catalog/) | Golden shape dictionary — copy `ex_*.json` nodes into the author dir; not a deliverable ([index](catalog/README.md)) |
© suzheng · Apache-2.0 · suzheng/k2f