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)

Letters are short works. The Orchestrator plans directly — no subagents for conception.

1.1 Identify the Emotional Core

One sentence: what must be communicated above all else? Not the topic — the feeling. Examples: love ("I need you to know what you changed in me"), condolence ("Your grief is witnessed"), apology ("I understand the specific harm I caused").

1.2 Calibrate the Relationship

Record: intimacy (strangers / acquaintances / friends / intimate / family), power dynamic (peer / sub-to-superior / superior-to-sub / mentor-mentee), history (long / recent / first contact), current state (strong / strained / broken / reconnecting).

1.3 Determine Voice

The letter must sound like the sender, not the writer.

If writing samples provided — analyze and extract:

1. Tone: formal vs. casual, warm vs. reserved, direct vs. circumspect
2. Cadence: sentence length patterns, fragments vs. complete sentences
3. Vocabulary: characteristic words, pet expressions, verbal tics ("anyway," "the thing is")
4. Emotional expression: direct ("I miss you") or indirect ("The house is quiet without you")?
5. Structural habits: long blocks? Staccato? Postscripts? Circling back?

Write a 2-3 sentence voice profile with 5+ distinctive markers.

If no samples — construct voice from context (age, background, relationship, occasion).

1.4 Save Plan

For a single letter: save a brief plan (emotional core, voice profile, relationship calibration, arc). No formal outline needed.

For letter series or epistolary collections: save full outline to /mnt/agents/output/{filename}.agent.outline.md.

Stage 2: Writing — Single Subagent (mandatory)

A letter is one person's voice. One writer. Never split.

For letter collections: strictly one letter per task call.

Create letter_writer

create_subagent:

• name: "letter_writer"
• system_prompt: Inline all of the following — (1) Sender voice profile (from Stage 1): the 2-3 sentence profile plus all markers. If from samples, include specific examples of the sender's phrasing. (2) Relationship context: intimacy, power dynamic, history, current state. (3) Anti-AI rules — read ../anti-ai.md and inline full content. For letters, especially emphasize: no template language ("I wanted to take a moment to express..."), no generic superlatives (show WHY, don't claim), specificity required (concrete memories beat "I cherish our time"), no copula avoidance ("It was good to see you" not "Seeing you brought warmth"), no rule-of-three forcing, imperfection is a feature (asides, unresolved feelings). (4) Letter conventions: salutation and closing carry meaning. 亲爱的 vs. 尊敬的; Dear vs. Dearest. Choose deliberately, not by template.
• description: "Write letter in sender's authentic voice"

Dispatch via task

task:

• agent: "letter_writer"
• prompt content:
• The emotional core (one sentence)
• Occasion and context
• Specific details to include: shared memories, inside references, concrete moments the sender would remember
• Length guidance (letters are rarely > 1500 words; most are 300-800)
• Format: date / salutation / body / closing / signature (adjust for medium)
• Medium conventions (if email: subject line, front-load purpose, shorter paragraphs, sign-off carries meaning)
• Output path: /mnt/agents/output/{filename}_sec01.md

After each letter completes, orchestrator runs word count verification:

python skills/general-writing/scripts/check_wordcount.py <file> --min {target} --lang auto

If FAIL, return to writer with expansion instructions.

Stage 3: Review

Step 1: Shared Pipeline

Read ../review.md. For a single letter, skip continuity_editor. Run style_editor (AI-pattern scan, voice consistency, rhythm) and structural_editor (arc: opening connection → core → closing feeling).

Step 2: Genre-Specific Editors

tone_editor

create_subagent:

• name: "tone_editor"
• system_prompt: "You are a tone editor for letters. Check: (1) occasion appropriateness, (2) formality match for the relationship, (3) emotional calibration — overwritten or underwritten? (4) medium-appropriate tone. Output: PASS/FAIL + specific feedback."
• description: "Tone and occasion appropriateness review"

task:

• agent: "tone_editor"
• prompt: Include letter file path, relationship calibration from Stage 1, occasion type, medium.

authenticity_editor

create_subagent:

• name: "authenticity_editor"
• system_prompt: "You are an authenticity editor for letters. Primary test: the template test — change sender and recipient names; if the letter still works, it is a TEMPLATE — FAIL. Also check: (1) specificity — 2-3+ concrete details, (2) AI-pattern scan — template openings, significance inflation, copula avoidance, rule-of-three, (3) imperfection — human texture present? (4) voice consistency through emotional shifts. Output: PASS/FAIL + template test result + specific feedback."
• description: "Authenticity and template-detection review"

task:

• agent: "authenticity_editor"
• prompt: Include letter file path, sender voice profile, anti-AI rules from ../anti-ai.md.

Scoring

Pass: >= 75 overall. Authenticity must be >= 70.

On FAIL: write 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 → re-review. Max 2 cycles. If still failing after 2 cycles, deliver with annotated issues for the user.

Stage 4: Delivery

Convert to .docx

Convert the final letter to Word using the md2docx pipeline (see skills/docx/SKILL.md → references/md2docx-reference.md).

Output: /mnt/agents/output/{filename}.agent.final.docx

This step is mandatory. Raw .md without a .docx deliverable is not an acceptable end state.

Occasion-Specific Notes

Reference Files