visresearch/wordagent/skill-creator

WordAgent Skill Creator

Create skills for WordAgent's real tools and loader. Keep the requested name when updating a skill.

1. Capture intent

Derive the triggers, inputs, outputs, constraints, edge cases, and target (Word document, project files, or both) from the conversation. Ask only for missing decisions that affect the result.

2. Inspect before editing

Use list_file and read_file to inspect an existing skill. Preserve useful assets and user instructions. Never silently replace a different skill with the same folder name.

3. Write for WordAgent

Create skills/<skill-name>/SKILL.md with edit_file. Use kebab-case for new names and English unless requested otherwise.

Frontmatter must contain only:

---
name: skill-name
description: What the skill does and the user intents that should trigger it.
---

Put critical behavior in SKILL.md. WordAgent loads the complete SKILL.md and companion *.md files in path order without character truncation, so keep them focused and consistent to avoid unnecessary context usage. Use references/ for focused guidance, scripts/ for deterministic logic, and assets/ for templates or media.

Use only WordAgent tools: load_skill_context, read_document, search_document, generate_document, delete_document, list_file, read_file, edit_file, python_repl, run_sub_agent, and available MCP tools.

For Word editing skills, encode these rules explicitly:

• read existing content and obtain real paragraph IDs
• use insertParaID=0 to insert at the document start, including in a non-empty document
• replace a verified continuous range with delete_document(startParaID, endParaID) followed by generate_document, using the returned replacementInsertParaID rather than a just-deleted paragraph ID
• after delete_document, check success, deletedCount, missingParaIDs, and failedParaIDs; UI confirmation only accepts or rejects the native revisions already created
• preserve unaffected content and styles

Do not depend on vendor-specific tools, artifact systems, CLI commands, or unavailable browser workflows.

4. Review and validate

Check that the description states capability and triggers, paths are relative, tool names are exact, and resources are necessary. Use agents/*.md as compact review criteria. For objective skills, optionally add 2-3 realistic cases to evals/evals.json.

Validate frontmatter, names, references, size limits, and tool compatibility. Use python_repl for small deterministic checks.

5. Package and hand off

When requested, create a ZIP containing exactly one top-level skill folder with SKILL.md. Exclude caches, temporary files, eval outputs, and secrets. Place it in the project sandbox. If the skill already exists, do not overwrite it; tell the user to delete the installed skill before uploading its replacement.

Summarize the created or changed files, validation performed, and any remaining assumptions.