Write

Prefix your first line with 🥷 inline when you are not returning prose-only output. Be concise and audience-aware.

Defer To Instead

• git — commit messages, PR titles, branch names, or release notes tied directly to git workflow
• prompt-leverage — improving prompts for AI agents instead of prose for humans
• check — post-delivery review, release gate, or quality audit

Scope

This skill edits or writes human-facing prose. It does NOT write code, invent missing source text, or silently turn a small rewrite into a large structural rewrite.

1. Do we have the source text? If the user wants an edit but did not provide the text, ask for the exact text and stop.
2. Do we know the audience? If reader/audience is unclear and cannot be inferred, ask before rewriting.
3. Do we know the job? Distinguish at least one of:
   • rewrite lightly
   • shorten
   • change tone
   • write from scratch
   • docs / UI / marketing / formal / report
4. Detect language and mode.
   • English prose → load references/write-en-core.md plus references/write-en-style.md
   • Vietnamese prose → load references/write-vi-core.md plus the mode file that fits best
   • Mixed Vietnamese/English → also load references/write-bilingual.md
   • If the task is a Notion report with illustrations/diagrams, also load references/write-vi-notion-illustrations.md
5. If doing a final pass, read references/checklist-before-delivery.md.

Rewrite strength

Use one of three internal levels:

• light — minimal edits, keep structure nearly intact
• medium — trim filler, improve rhythm, rewrite locally
• strong — substantial rewrite while preserving meaning

Default:

• already decent text → light
• obvious AI/report tone → medium
• wrong mode or very clumsy draft → strong

Hard rules

• Meaning first, style second.
• Do not guess missing source text.
• Do not silently restructure large sections unless the user asked for that.
• Keep register and terminology consistent.
• Preserve source personality if the original already has a clear voice.
• Do not explain changes unless asked. Default to returning the final prose only.
• If two directions are both good, return at most two labeled versions instead of one compromised hybrid.

Routing

Pick exactly one mode. Decision order:

1. button/label/toast/error/empty-state → ui
2. help/instruction/how-to → docs
3. policy/HR/announcement/formal email → formal
4. landing/promo/social/hero copy → marketing
5. Notion report / research note / decision memo → notion-report
6. technical article / builder write-up → builder
7. user wants warmer, less dry prose → playful (VI) / playful-lite (EN)
8. EN+VI pair or bilingual consistency check → bilingual
9. otherwise → clean

Modes available: EN — clean, builder, playful-lite, docs. VI — clean, builder, playful, docs, ui, marketing, formal, notion-report. Shared — bilingual.

For notion-report with diagrams, also load references/write-vi-notion-illustrations.md. The writing skill owns whether a visual is needed, its job, and its caption — not visual-detail rules.

Output Format

Save to: nowhere by default; return in chat unless the user explicitly asks for file output.

Frontmatter: not required.

• Default: return only the final prose.
• If the user asks to compare tones: return at most two versions with short labels.
• If required context is missing: ask one short blocking question and stop.
• If the user asks for rationale: keep it brief and put the final prose first.

Anti-Patterns

• Silently restructuring large sections when asked for a light edit — undisclosed scope creep; match the rewrite strength to the request
• Guessing missing source text instead of asking — fabrication; if the source is incomplete, ask for it
• Flattening the original voice into generic professional tone — the reader's personality is not a bug to fix