personal-writing-voice

---
name: personal-writing-voice
description: Chinese influence-research and personal writing-transfer system with V5 source screening, four-layer language and argument fingerprints, evidence-backed transfer plans, author/material/style gates, close-imitation review, reader simulation, feedback learning, multi-style isolation, and private long-term memory. Use when Codex needs to help a writer think, research Chinese original articles, plan, draft, revise, de-AI, review, publish, or learn from feedback while preserving the writer's own judgment and voice.
---

# Personal Writing Voice V5

Use this skill as a Chinese influence-research and writing-transfer system. Keep the author's real judgment and material in control while using public articles only as evidence for expression mechanisms.

Default to the user's language. For Chinese prompts, answer in Chinese unless asked otherwise.

## Fixed Priority

Apply this order without exception:

```text
author's real judgment
> author's real material
> current task need
> active personal style
> external expression mechanisms
> general writing technique
```

Never let a reference article decide what the author should think or say.

## Core Rules

- Select one active style before substantial drafting or revision. Keep author facts separate from style rules and task influence.
- Establish author state, core judgment, material ledger, factual gaps, emotional anchors, light-touch material, and do-not-invent boundaries before prose.
- Ask at most three high-value questions at a time. Keep internal rigor out of the user's way.
- Use Chinese-original sources only for Chinese wording, syntax, and rhythm. Use translations or foreign-language sources for facts or ideas only.
- Borrow mechanisms, not finished expression. Never copy signature wording, metaphors, recognizable cadence, transition sequences, or a full article structure.
- Require locatable short evidence, suitability, use position, intensity, and imitation risk for every transferable feature.
- Keep influence task-specific. Do not promote a source feature into permanent memory without explicit author confirmation.
- Preserve useful irregularity, hesitation, restraint, local phrasing, and uneven rhythm when they carry the author's real stance.
- Separate deterministic warnings, editorial judgment, reader experience, and memory-gate results. Do not produce a false total quality score.
- Protect private profiles, drafts, materials, feedback, source notes, task records, and context packs from public export.

## Mode Routing

Recognize the user's natural commands:

- `快速写` → `fast`
- `正常写` → `standard`
- `准备发表` → `publish`

### Fast

Use for short writing, local polish, and low-risk work.

- Do not require external influence research.
- Build only the material boundaries needed for the task.
- Check author voice, factual risk, language naturalness, and unauthorized rewriting.

### Standard

Use for ordinary articles.

- Select 1-2 screened sources.
- Transfer 3-4 evidence-backed features.
- Run material, source, style, transfer, and close-imitation checks.

### Publish

Use for public-account long-form, important speeches, and formal submissions.

- Select 3-5 sources with distinct roles.
- Include at least one verified Chinese-original source published on or before 2016-12-31.
- Transfer only 3-6 evidence-backed features.
- Complete full material, fact, style, source, transfer, editorial, reader, privacy, and memory gates.

If the user forbids browsing and publish mode cannot satisfy its source gate, stop before prose and report the blocker.

## Required Workflow

1. Understand why the author writes now, what they truly judge, what remains uncertain, what must not be written, and what change the reader should experience.
2. Select the active style and compile author-level, active-style, and task rules without cross-style leakage.
3. Build the material ledger. Do not invent experience, emotion, motivation, causality, facts, or conclusions.
4. Route by mode. For standard/publish, search, screen, and classify sources before prose.
5. Analyze selected sources across lexical choice, sentence generation, argument progression, and article rhythm.
6. Build influence cards and a task transfer plan. Keep only 3-6 useful features; do not allow multiple high-intensity source controllers.
7. Run `draft-preflight`. Do not draft when a required source, feature, material, evidence, style, or timing gate fails.
8. Write the thought draft first: truthful judgment, sound material sequence, paragraph progression, and no content distortion for resemblance.
9. Apply language transfer only after the thought draft is stable.
10. Review author voice, transfer outcomes, facts, close imitation, editorial value, and reader experience.
11. Run the rule memory gate on the stable candidate. Make only authorized local revisions and verify again.
12. Store feedback as candidate evidence. Promote, conflict-mark, or retire rules only through explicit memory decisions.

## Reference Routing

Read only the references required for the task:

- Read `references/thinking-and-materials.md` for author thinking, material ledgers, factual gaps, and boundaries.
- Read `references/style-library.md` for style selection, creation, storage, and isolation.
- Read `references/style-blending.md` only for explicitly requested blends.
- Read `references/chinese-source-screening.md` before selecting any language or publish-mode influence source.
- Read `references/chinese-influence-analysis.md` before extracting source features.
- Read `references/influence-transfer.md` before building influence cards or transfer plans.
- Read `references/influence-safety.md` before drafting from sources or checking close imitation.
- Read `references/influence-research.md` for the complete search-to-record workflow.
- Read `references/writing-workflow.md` for task sequencing, thought drafting, language transfer, and revision.
- Read `references/editorial-review.md` for author-voice, transfer-outcome, and article-worth judgment.
- Read `references/evaluation.md` for deterministic, editorial, reader, and acceptance checks.
- Read `references/reader-simulation.md` for publish-mode or important long-form reader checks.
- Read `references/memory-gate.md` before preparing or verifying final rule coverage.
- Read `references/revision-safety.md` before copyediting, line editing, de-AI work, or any rewrite.
- Read `references/feedback-loop.md` before learning from author feedback.
- Read `references/failure-recovery.md` when the author says “不像我” or rejects the draft's meaning.
- Read `references/privacy-and-storage.md` before writing private files, context packs, source notes, or exports.
- Read `references/author-intake.md`, `references/sample-analysis.md`, and `references/author-profile-template.md` when establishing or updating a voice library.
- Read `references/anti-ai-review.md` for de-AI requests.
- Read `references/command-protocol.md` for compact commands and CLI mapping.
- Read `references/examples.md` only when examples are needed.

## Hard Gates

Do not pass when any applicable condition fails:

- A language source is not verified `zh_original`.
- A selected source is an untraceable repost, aggregation, suspected SEO farm, scraped copy, or suspected batch-AI article.
- A transferable feature lacks short evidence, a locator, intended use, intensity, or risk.
- `do_not_borrow` is empty for a selected standard/publish source.
- A task feature belongs to another style.
- More than one source controls the task at high intensity.
- Required research or `draft-preflight` occurs after the draft was created.
- Material is missing, unverified, forbidden to invent, or unsupported where the claim requires it.
- A task-level feature is promoted into permanent memory without author confirmation.
- A revision exceeds its authorized mode or changes fact, intent, experience, emotion, or judgment without permission.

## CLI

Use `scripts/pwv.py` for file-backed workflows. The three primary entry points are:

```text
python scripts/pwv.py prepare-task --author cc --style wechat-reflection --task-id task-1 --mode publish --title "题目"
python scripts/pwv.py check-draft --author cc --style wechat-reflection --task-id task-1 --draft draft.md
python scripts/pwv.py learn-feedback --author cc --style wechat-reflection --feedback "这段不像我"
```

Use low-level compatible commands when needed:

- `add-task-influence`, `draft-preflight`, `migrate-influence`
- `add-material`, `check-materials`
- `prepare-gate`, `verify-gate`, `gate-report`
- `add-feedback`, `distill-feedback`, `promote-rule`, `retire-rule`
- `compare-revision`, `evaluate-draft`, `reader-sim-template`
- profile, privacy, migration, validation, and export commands

The CLI never browses or calls an LLM. The agent performs source research and semantic/editorial judgment; the CLI records evidence and enforces deterministic gates.

## Delivery

Default to a one-screen user report:

1. Core judgment understood
2. Missing or risky material
3. Active style and mode
4. Selected transferable features
5. Decisions still required from the author

For final prose, deliver the body first, then a one-line gate summary. Keep detailed source cards, transfer records, mechanical warnings, and full rule reports private unless requested.