Skills

Skills are reusable procedures you make available from the chat: a review checklist, a brand-voice guide, a data-cleaning workflow. They follow the open Agent Skills model: a skill's name and description are advertised to the model, and its instructions load only when they are actually needed.

Skills are not always-on rules. Where Custom Instructions apply to every turn, a skill is opt-in and loaded on demand.

Availability per app

A skill is available only in the Office apps you choose, so a Word skill never clutters Excel and vice versa. In the skill editor, tick the apps where the skill applies: Word, Excel or PowerPoint. New skills default to all three.

In the skills panel, each skill shows its app badges (W X P); skills that don't apply to the app you're in are dimmed. The model only ever sees the skills available in the current app.

Using a skill

A skill's text is not loaded just because it's available; only its name and description are advertised, so the model knows it exists:

1. Discovery: the model sees the name and description of every skill available in the current app.
2. Activation: when your request matches a skill's description, the model loads its full instructions (via the load_skill tool) and follows them. You can also invoke a skill yourself by typing / in the input and picking it (or typing /slug). Text after the command becomes the request.
3. Execution: the model carries out the procedure using its normal tools, including execute_office_js.

Because only the description is kept in context until a skill is used, you can keep a large library with a small context footprint.

Writing a skill

Field	Required	Notes
Name	yes	Shown in the menu, up to 60 characters. Also drives the slug used to invoke it (e.g. Concise answers → /concise-answers).
Description	yes	What it does and when to use it (up to 1,024 characters). This is the activation hint the model matches against, so be specific.
Instructions	yes	The procedure loaded when the skill is activated (up to 8,000 characters).
Available in	yes	Word, Excel and/or PowerPoint (at least one).

Office.js snippets as reference

A skill's instructions are plain text, so you can include Office.js code snippets for the model to use as reference, for example a batch of formatting calls or a known-good pattern to adapt. The model can run them through execute_office_js . This is the closest equivalent to the standard's bundled scripts, which OpenDocBot does not execute.

Sharing skills

Skills use the open Agent Skills format, so they work across tools (Claude Code, Cursor, GitHub Copilot, and more):

• Export skills downloads opendocbot-skills.zip, one <slug>/SKILL.md entry per skill.
• Import skills reads a .zip pack or a single SKILL.md and merges it into your library by slug.

A SKILL.md is YAML frontmatter plus a Markdown body. The app scope is stored in metadata.hosts (a comma-separated list, always written; if absent on import the skill is available in every app):

---
name: concise-answers
description: Short answers for busy readers.
metadata:
  display-name: Concise answers
  hosts: word,excel
---

Be concise. Prefer short sentences and bullet points.

Skills are instructions

A skill's instructions are followed by the model with the same trust as your own. Only import skills from people you trust, and review a skill before making it available.

Differences from the full standard

OpenDocBot runs in the Office task pane, without a filesystem or a code sandbox. Skills here are therefore instructions only: the optional scripts/, references/, and assets/ of the full specification are not executed or loaded. Work that the standard would put in a script is done through the document tools instead (for example execute_office_js).

Skills are stored in your browser only, alongside the rest of your settings; no skill content is sent anywhere until it is activated in a message.