letter

Letter Writing

Orchestrator dispatches all creation and review work to subagents. Before creation, copy user-uploaded files to /mnt/agents/output/research/.

Sub-skill of general-writing. Inherits shared creative writing principles, anti-AI discipline, and quality gate framework.

CRITICAL — Review → Fix Brief → Dispatched Subagent: When a review finds issues, translate them into a detailed fix brief (file paths, quoted findings, expected outcome, scope boundary, self-check) and dispatch a fresh letter_writer subagent to apply the fix. Never use inline sed/edit_file fixes.

CRITICAL — Anti-Read-Loop Rule: The orchestrator must dispatch its first subagent (letter_writer) within 10 iterations. If the orchestrator has spent 10+ iterations only reading files without creating any subagent, it is in a read-loop. Stop reading, draft your conception from what you have, and dispatch the writer immediately.

CRITICAL — .docx Delivery: Every completed letter must be converted to .docx using skills/docx/SKILL.md. Never end a session with raw .md files and no .docx deliverable.

File Paths

Output: /mnt/agents/output/
  {project_name}.letter.md              # Final letter
  research/                              # User uploads (writing samples, context docs)

Trigger Keywords

书信, letter, 情书, 公开信, 遗书, 感谢信, open letter, epistle, 致辞, 家书, 道歉信, farewell letter, condolence letter, 推荐信, cover letter, application letter, email, 求职信, 申请信

---

Pre-Stage: Context Inference

Infer letter type, sender/recipient relationship, occasion, tone, and length from the user's request. The orchestrator analyzes and infers all parameters autonomously; for complex requests, deploy a subagent to analyze requirements in depth. Never prompt the user for clarification.

If the user provides writing samples from the sender — collect them. Voice adaptation is mandatory when samples exist.

---

Stage 1: Conception (Orchestrator — no subagents)