← Spell Book

tw-formal-writing

ai-tools

Spell Rating
🔮🔮🔮🔮○
Pricing
free
Difficulty
intermediate

Best For

Anyone who needs to write formal Taiwan government documents (official letters, memos, proclamations) or legal documents (certified letters, contracts, NDAs) and does not have the formatting conventions memorized. This is a niche tool for a niche problem: LLMs have near-zero awareness of Taiwan's official document conventions, and getting the structure wrong means your letter gets returned or ignored.

If you work in Taiwan's public sector, academia interfacing with government agencies, or legal practice involving standard-form documents, this is directly relevant. If you never write documents to or within Taiwan's government system, this tool has no use for you.

How I Actually Use It

I keep it on watch status. My current need for formal government correspondence is infrequent, but when it comes up (grant application cover letters to NSTC, administrative memos within the university), I plan to install the SKILL.md version into Claude Code and use it for first-draft generation. The more practical mode is "review and fix": paste an existing document and get a list of formatting and terminology errors. That is more defensible than generating from scratch, because the human wrote the content and the AI only checks the form.

Installation is a single git clone into ~/.claude/skills/. No dependencies, no runtime, no API keys.

Where It Is Strong

Knowledge routing pattern. SKILL.md does not dump 26K characters of rules into the context. It acts as a router, loading only the relevant reference file (official-letter.md, legal-documents.md, etc.) based on the user's request. This is lightweight RAG without a vector database -- just file structure. It is a design pattern worth studying for anyone building AI Skills.

Three deployment versions. SKILL.md (multi-file, Claude Code), STANDALONE.md (single file, ChatGPT/Gemini), LITE.md (2K chars, GPTs Instructions). Each version makes a deliberate trade-off between knowledge completeness and platform compatibility. This is engineering discipline, not a copy-pasted prompt.

Review and fix mode. You can paste an existing document and get an error list against the official regulations. This is arguably more useful than generation, because it catches mistakes in documents you already wrote.

Solid legal references. Not vague "follow official standards" -- specific citations to the 112-year edition of the Document Processing Manual, the Official Document Format Act, and administrative interpretations updated to 115.2.23.

Built-in quality gate. Every generation automatically runs a four-dimension checklist (structure, terminology, character usage, format compliance) without the user needing to ask.

Where It Fails

No independent validator. The quality gate is the same LLM checking its own output. When the model hallucinates a nonexistent regulation article number, the built-in checklist will not catch it. For legal documents, this is a structural weakness.

Knowledge ages. Regulations get amended, administrative interpretations get issued. The knowledge base is static Markdown. After v1.1.0, any regulatory changes require the maintainer to manually update the reference files.

No programmatic citation verification. Legal article numbers and interpretation reference numbers are hardcoded strings. There is no connection to any regulatory database to verify they exist or are current.

Single maintainer. Cheng-I Wu (Imbad0202) maintains this alone. If maintenance stops, the knowledge base freezes at whatever regulatory version was last updated.

Pricing / Difficulty / Risk

  • Pricing: Free, MIT license.
  • Difficulty: Intermediate. You need to understand how Claude Code Skills work (or know how to paste STANDALONE.md into ChatGPT). The actual document generation is guided, but you need enough domain knowledge to verify the output.
  • Risk: Low for drafting, moderate for final submission. Never submit a generated document without human review. The tool helps you get the structure right, but it cannot guarantee legal accuracy.

Verdict

Rating 4. Watch status. The problem definition is precise and the solution architecture is thoughtful. The knowledge routing pattern and three-version strategy show real engineering care. But for a domain where formatting errors have real consequences, the lack of independent verification is a hard ceiling on trust. Use it for first drafts and review passes, not as a final authority.

Source