create-double

name

create-double

description

Build and maintain a local, private digital double from interview answers and freeform self-description. Use when Codex needs to ask one high-value self-modeling question at a time, accept unstructured personal description, merge explicit corrections, and generate `profile.yaml`, `profile.md`, and a runnable double `SKILL.md` inside the local `doubles/slug-name/` folder without inventing memories or overstating certainty.

Create Double

Overview

Build a Chinese-first digital double that stays local to this repository. Keep profile.yaml as the main runtime record, use session.yaml for transient interview state, and regenerate profile.md plus the runtime SKILL.md from the profile.

The repo now also has a private knowledge-base layer:

• doubles/<slug>/kb/ high-signal long-term knowledge for one double
• .project-kb/ private maintainer knowledge for this repo

That layer is for accumulation and evidence. It does not replace profile.yaml as the runtime source of truth.

Quick Start

Fast path for a human first run

If the user only wants the first visible artifact, prefer:

python scripts/double_builder.py start --slug my-work-double --display-name "工作分身" --use-case work

If the user wants to check the environment first, use:

python scripts/double_builder.py doctor
python scripts/double_builder.py start --demo --use-case work

start now has two front-door decisions:

• --use-case general, work, self-dialogue, external, or custom
• --depth quick, standard, or deep

Recommended defaults:

• quick 3 base questions, render first artifact, one natural-language correction, then optionally continue with 2 more questions
• standard 3 base questions + 2 follow-up questions
• deep 3 base questions + 4 follow-up questions, including at least one anchor example

Use correct right after generation when the user says “我不会这么说” or “我更在意 X”.

python scripts/double_builder.py correct --slug my-work-double

start and correct now also capture private high-signal knowledge events under doubles/<slug>/kb/. Keep that automatic capture lightweight and local-only.

Low-level path for Codex and custom workflows

1. Initialize a new double:

   python scripts/double_builder.py init --slug my-double --display-name "我的分身"
   

2. Route the latest user turn:

   python scripts/double_builder.py route --current-mode interview --text "我做决定时通常先看长期影响"
   

3. Read only the prompt file you need:

   • prompts/router.md for turn classification
   • prompts/interview.md for use-case-aware question selection
   • prompts/freeform.md for extracting structured data from self-description
   • prompts/correction.md for handling repair turns
   • prompts/rendering.md before using the generated double

4. Convert the user turn into a structured payload that follows references/payloads.md.

5. Apply the payload and render outputs:

   python scripts/double_builder.py apply-turn --slug my-double --payload-file payload.json
   python scripts/double_builder.py render --slug my-double
   

6. Continue interviewing or answer through the generated double in doubles/my-double/SKILL.md.

Core Rules

• Treat 继续提问, 我自己说, 先生成看看, and 我要改一下 as hard control phrases.
• Ask at most one question at a time. Prioritize values, decision tradeoffs, voice, and boundary style before biography details.
• Keep direct user statements and model inferences separate. Use source: direct only when the user explicitly said it, source: inferred only for tentative synthesis, and corrections to override previous guesses.
• Never fabricate memories, dates, relationships, or preferences that are not anchored in the profile.
• If the profile is sparse, say so plainly and label any advice as inference.
• Keep every artifact local to doubles/<slug>/. Do not depend on external services.
• Treat each double as single-use-case-first. If the user wants a work self, a self-dialogue self, and an external-facing self, prefer separate slugs over one overloaded profile.

Workflow

1. Initialize or Resume

• Create a slug in lowercase hyphen-case.
• Run python scripts/double_builder.py init --slug <slug> --display-name "<name>".
• Reuse the existing folder if the double already exists. Read profile.yaml, profile.md, and session.yaml before asking new questions.

2. Route Every Turn

• Use python scripts/double_builder.py route --current-mode <mode> --text "<user-turn>" as a first-pass classifier.
• Confirm or override the route with your own judgment.
• Map each turn to one of answer, freeform, correction, switch_mode, or finish.
• Prefer correction when the user is refining prior language or values. Prefer freeform when the user gives a dense self-description. Prefer answer when the user is clearly answering the last question.

3. Build a Structured Payload

• Read references/profile-schema.md if you need field definitions.
• Read references/payloads.md for the JSON shape expected by apply-turn.
• Fill only the supported slots and keep the update small.
• Refresh unknowns with the top remaining gaps and set next_question to the single best follow-up.
• Keep meta.primary_use_case, session.interview_track, session.interview_depth, session.pending_questions, and session.asked_questions aligned with the active interview track.

4. Apply and Render

• Apply the payload with apply-turn.
• Render when the user says 先生成看看, when a meaningful amount of new information has landed, or before you want to answer as the generated double.
• Rendering must preserve a snapshot under history/ before overwriting prior generated artifacts.

5. Use the Generated Double

• The generated double is for two jobs only:
  • Talk like this person would talk.
  • Give advice using this person's values and decision model.
• If asked about experiences that are not present in anchor_examples, state that the answer is incomplete or inferred.
• If the user later corrects the double, write the correction to corrections, update the relevant fields, and render again.

Adaptive Interview Notes

• general Focus on decisions, advice style, and boundaries.
• work Focus on collaboration style, risk tradeoffs, and expectation management.
• self-dialogue Focus on self-talk, self-deception boundaries, and what helps restore clarity.
• external Focus on outward expression, impression, and public or semi-public boundaries.
• custom First ask what the user wants the double to help with, then map to the nearest track.

Output Contract

• Canonical profile: doubles/<slug>/profile.yaml
• Human-readable profile: doubles/<slug>/profile.md
• Generated runtime skill: doubles/<slug>/SKILL.md
• Transient session state: doubles/<slug>/session.yaml
• Private long-term knowledge: doubles/<slug>/kb/
• Snapshot history: doubles/<slug>/history/<timestamp>__v<version>/

Knowledge Base Contract

• Treat profile.yaml as the only runtime structured truth.
• Treat doubles/<slug>/kb/ as a private accumulation layer for stable evidence, corrections, and longer-lived context.
• Do not store full life-log transcripts by default.
• Do not invent knowledge pages that imply certainty beyond what the user actually said.
• If a knowledge event is still unstable or conflicting, leave it in the KB or unknowns instead of forcing it into profile.yaml.

When To Read Extra Files

• Read prompts/router.md only when the route is ambiguous.
• Read prompts/interview.md only when choosing the next question.
• Read prompts/freeform.md only when extracting structured data from long self-description.
• Read prompts/correction.md only when a turn is mainly a repair.
• Read prompts/rendering.md only when generating or auditing the final double behavior.