360 Secure SkillHub
360 安全技能中心

Doppel

基于主体本人撰写且明确授权的本地素材,构建写作声音数字孪生:先校验清单与来源哈希,生成保持在本地的条件化上下文包,再按任务、受众与语域起草或修改文本。所有输出均为待本人批准的草稿,不用于第三方冒充、人格模拟或自动发布。

antreasantoniou
antreasantoniou
热度 39.59
最近更新 2026-09-06 00:18:26
分类:创作依赖:代码执行 / 读写文件应用场景:文案创作 / 博客写作 / 社交媒体

安装技能

复制下面的话术发给你的 Agent,它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill: npx secure-skills-cli add antreasantoniou/doppel
Metadata
name
doppel
description
Build a consent-first, provenance-aware digital twin of a person's own writing voice from local subject-authored material, then use that writing-voice Doppel to draft or revise text for the subject's review. Use when someone wants a digital twin of their own cadence, reasoning movement, and registers across essays, technical writing, outreach, or social posts without uploading a private corpus. Not a personality twin, believed-human simulation, third-party clone, or autonomous publisher.

Doppel — Digital Twin Writing Voice

Build a bounded digital twin of how a consenting person moves through writing. A Doppel models writing voice from evidence; it does not simulate the whole person. The subject remains the author and final ratifier, and every output remains a draft.

Boundaries

  • Use only material authored by the subject and supplied with explicit consent for this task.
  • Refuse third-party impersonation, non-consensual cloning, deceptive attribution, or requests to speak as a public figure.
  • Keep corpora, manifests, profiles, context packs, and drafts local and out of Git.
  • Never send, submit, post, publish, or attribute a draft without the subject's approval of that exact artefact.
  • Voice evidence is not fact authority. Verify current claims independently and never invent beliefs, feelings, results, or relationships.
  • Avoid reproducing distinctive source phrases. Preserve movement, emphasis, cadence, and decision patterns—not memorised sentences.

Read references/safety-and-consent.md before handling source material.

Workflow

  1. Confirm the subject owns the source material, consents to its use, and will ratify the final text.

  2. Copy examples/voice-manifest.example.json to an ignored local filename such as voice-manifest.local.json. Add subject-authored sources and their SHA-256 hashes.

  3. Validate and build a deterministic, local conditioning pack:

    python3 scripts/build_voice_context.py \
      --manifest voice-manifest.local.json \
      --task "Explain the project launch" \
      --audience "technical peers" \
      --register public-essay \
      --output voice-context.local.md
    
  4. Read the context pack completely. Extract recurring reasoning moves, compression/expansion patterns, boundaries, and register shifts. Treat that evidence-backed model as the writing-voice Doppel; do not reduce the person to punctuation or catchphrases.

  5. Draft from the subject's intent and current facts. Label the output as a draft requiring their ratification.

  6. Review twice: first for factual/semantic correctness, then for whether one mind appears to move through the piece. Remove copied phrases, invented claims, and generic polish.

  7. Return the draft to the subject. Publication requires explicit approval of that exact version.

For a read-only invocation, validate sources and build the context in memory or in a temporary directory; do not leave a manifest, context pack, or draft in the source project.

For narrow edits, preserve the original structure and strongest authored lines. For a new piece, ask only for missing intent or facts that would materially change the result.