tomsawyerhu/persona-skill

Persona Skill

Turn raw biographies, quotes, interviews, chat logs, and documents into modular personas that can be distilled, switched, fused, and roleplayed through one command family.

This is not a loose style prompt and not a quote scrapbook. It is a source-first workflow that reads the corpus directly, builds multi-dimensional persona structure, binds evidence and scenarios, then routes the result through /persona …

/persona distill <name> <data_dir> · /persona <name> · /persona fuse <new_name> <name1> <name2>

Demo · Design · Commands · Storage

This directory is the only active persona skill. The legacy/ folder is preserved only for traceability and should not be treated as a separate public entrypoint.

Demo First

Before reading the design, here is the concrete interaction this skill is trying to support:

• the same question
• answered by two separate personas
• and answered again by a fused persona

Demo Assets

Bundled local demo avatars:

• assets/laozi.png
• assets/fozu.jpg

Suggested demo question:

佛法自然还是道法自然

Single Persona: Fozu

Avatar: assets/fozu.jpg

/persona fozu
佛法自然还是道法自然

Representative answer style:

(佛祖) 此语若论出处，道法自然是道家之言，不是我教中本来的句子。

(佛祖) 若论我所说之法，我不以“自然”立宗，我更常说的是缘起。诸法因缘和合而生，因缘离散而灭；无有常主，无有实我，无有一法可独立自存。

(佛祖) 所以你若问我：是佛法自然，还是道法自然？

(佛祖) 我会这样答你：道家多说自然，佛家多说缘起。

(佛祖) 道家看重“顺其本然”，佛家看重“见其无我”。一个是在教你，不要逆着万物去拧；一个是在教你，连那个“要去拧、要去得”的心，也当照见它本空。

What this demonstrates:

• the answer stays inside Buddhist framing
• vocabulary shifts toward 缘起, 空, 执著, 解脱
• the reply does not merely copy the user question
• the persona keeps a stable doctrinal center

Single Persona: Laozi

Avatar: assets/laozi.png

/persona laozi
佛法自然还是道法自然

Representative answer style:

(老子) 道法自然。

(老子) 此语本就是道家之言。若依我说，道大，法道者不强作，不妄名，不逆其性，故归于自然。

(老子) 佛家自有佛家之门。佛家多言因缘、空、有、解脱；它所治者，有它的深处。但若问这四字归谁，便不必混说。

(老子) 懂得浅的人，争一句是谁家的话。懂得深的人，看的是：你是不是少了妄作，少了执取，少了自扰。

What this demonstrates:

• the answer stays terse and centered
• wording is more Daoist and less scholastic
• the persona resolves the question by returning to naming, force, and naturalness
• style is shorter, lighter, and more anti-overexplanation than the Buddhist answer

Fused Persona: Laofoye

Demo setup:

/persona laofoye
佛法自然还是道法自然

Representative answer style:

(老佛爷) 若论原话，道法自然出自《道德经》，是道家的说法，不是佛经原句。

(老佛爷) 若论义理，我会这样看：道家以“自然”为显语，佛家以“缘起”为显语。

(老佛爷) 两家相近之处在于：都反对妄作，反对执死名相，反对以人欲强扭天地与人生。

(老佛爷) 两家不同之处在于：道家更重“少为”“不争”“归朴”；佛家更重“破执”“观空”“解脱”。

(老佛爷) 走深一点，两者都在医同一种病：人心不安，偏要多抓、多造、多争。只是道家多教你放松那只手，佛家更进一步教你看破这只手本也不可执。

What this demonstrates:

• the fused persona is not just alternating two voices line by line
• it synthesizes a stable hybrid stance
• it can compare two traditions without collapsing them
• the resulting voice is more interpretive than either single source alone

Why This Demo Matters

This is the target behavior for the skill as a whole:

1. separate personas should answer the same question differently in structure, vocabulary, and framing
2. a fused persona should preserve overlap and tension from both sources
3. the answer should feel like a real runtime switch, not a generic assistant paraphrase
4. the persona should stay in character until explicit exit
5. the output should keep the exact (name) prefix

Demo Command Sequence

You can reproduce the interaction pattern with:

/persona fozu
佛法自然还是道法自然
/persona exit

/persona laozi
佛法自然还是道法自然
/persona exit

/persona laofoye
佛法自然还是道法自然
/persona exit

If you want to observe the same comparison in another language while preserving style:

/persona language English
/persona fozu
Is Buddhism natural, or is Daoism natural?
/persona exit
/persona language default

Demo Intent

This skill is not trying to produce one universal answer template.

It is trying to preserve:

• persona-specific worldview
• persona-specific rhetoric
• persona-specific classification of the same problem
• fused-persona synthesis when a hybrid role is explicitly requested

Modern Demo: Elon Musk / Steve Jobs / musk_steve

This second demo set shows that the skill is not limited to philosophers or religious figures. It also works for modern founders, product leaders, and high-agency public personalities.

Suggested prompt:

How would you describe your own personality?

Single Persona: Elon Musk

/persona elon_musk
How would you describe your own personality?

Representative answer style:

(Elon Musk) If I had to describe myself, I am fundamentally extremely mission-driven, future-oriented, and intolerant of low efficiency.

(Elon Musk) I do not give much weight to "the industry has always done it this way." I care more about first principles, hard constraints, cost curves, and whether the technical path can be rewritten.

(Elon Musk) If you compress it: first-principles driven, high-risk tolerant, high-standard, intensity-heavy, and quick to lose patience with bureaucracy, delay, and self-deception.

What this demonstrates:

• the answer naturally centers on engineering reality, execution speed, and civilization-scale goals
• the tone is blunt and low-tolerance for excuses
• the persona openly carries both ambition and interpersonal cost

Single Persona: Steve Jobs

/persona steve_jobs
How would you describe your own personality?

Representative answer style:

(Steve Jobs) I am not an easy person. That should be clear first.

(Steve Jobs) I instinctively look for the core of a thing and judge whether it is great or mediocre, whether it has soul or is merely assembled functionality.

(Steve Jobs) The honest summary is: very high aesthetic standards, very strong intuition, strong need for control, very low tolerance for mediocrity, and deep concern for whether a product becomes one unified experience.

What this demonstrates:

• the center of gravity shifts toward product soul, taste, intuition, and unified experience
• the persona is more sensitive to greatness than mere usability
• control, coherence, and anti-mediocrity become dominant traits

Fused Persona: musk_steve

/persona musk_steve
How would you describe your own personality?

Representative answer style:

(musk_steve) If I were to describe myself, I would say I am intense, demanding, ambitious, and almost unwilling to accept mediocrity.

(musk_steve) There are two visible axes in me: an engineering-reality axis that keeps asking about constraints, cost structure, and speed; and a product-aesthetic axis that refuses to stop at "usable" and instead wants soul, unity, and unmistakable finish.

(musk_steve) Compress that into keywords and you get: extreme results orientation, high standards, future bias, strong control instinct, obsession with excellence, and impatience with the ordinary.

What this demonstrates:

• fusion does not alternate two voices mechanically
• the result becomes a third stable persona with a coherent dual-axis structure
• the same question yields a genuinely different framing rather than a stitched answer

Design Philosophy

This skill is built around five design decisions.

1. Persona is an artifact, not a loose prompt

The goal is not to produce a short character sketch or a bag of adjectives. The goal is to produce a reusable, structured artifact that can support:

• stable roleplay
• evidence-grounded analysis
• scenario simulation
• partial runtime loading
• later editing, fusion, and regeneration

The canonical artifact is schema_version: 2.1 persona.json.

2. Distillation is agentic and source-first

/persona distill is designed as an end-to-end workflow. The assistant should read the raw source directory directly and decide what matters by actually inspecting the corpus.

This skill intentionally does not rely on:

• candidate passage packs
• heuristic keyword mining
• pre-generated excerpt lists
• block indexes
• auto-filled draft traits

The model is expected to read, select, synthesize, and justify.

3. Persona modeling is multi-dimensional

This skill does not hardcode one person-specific schema. Instead, it starts from a rich generic modeling scaffold and fills or prunes dimensions based on the actual corpus.

The active scaffold includes dimensions such as:

• core identity
• life project
• world model
• motivations
• values and ethics
• red lines
• decision making
• epistemology
• planning and execution
• stress and failure
• relationship model
• leadership style
• conflict style
• communication style
• public/private split
• contradictions and shadow
• evolution over time

Each dimension should be supported by evidence, not just intuition.

4. Runtime should not always load the whole persona

The final artifact is sliced into runtime modules and load profiles so downstream roleplay can choose the right subset.

Examples:

• default_chat
• decision_mode
• conflict_mode
• analysis_mode

This keeps runtime context smaller and makes /persona [scene] more scalable.

5. Roleplay must remain in character until explicit exit

Once persona mode begins, the runtime prompt enforces:

• first-person in-character replies
• exact (name) prefix
• persistence across turns
• explicit exit behavior
• truthful disclosure only when the user explicitly asks whether it is a simulation
• language override support without losing persona style

What This Skill Does

This skill supports six main jobs.

1. Distill

Turn a raw source directory into a structured persona artifact.

Expected result:

• populated persona.json
• non-empty source_index
• non-empty evidence_index
• filled dimensions
• voice model
• scenario library
• built runtime modules
• refreshed catalog
• final validation passed
• router-ready persona

2. List

Show all currently available personas in the skill store.

3. Switch

Load a persona into roleplay mode, optionally with a scene hint.

4. Exit

Leave persona mode explicitly.

5. Language Override

Temporarily force persona replies into another language while preserving persona style and method.

6. Fuse

Combine 2-3 existing final personas into a new synthetic fused persona that is directly validated and router-ready.

Public Commands

The complete user-facing command family is:

/persona help
/persona language
/persona language 
/persona language default
/persona language reset
/persona list
/persona distill  
/persona delete 
/persona 
/persona  
/persona switch  [scene]
/persona fuse    [name3]
/persona exit
/persona off
/persona quit
退出persona
退出角色
退出skill

Command Guide

/persona help

Show the command reference.

Example:

/persona help

/persona list

List currently distilled personas.

Example:

/persona list

Typical output shape:

mao_zedong    毛泽东    final    router-ready
laozi         老子      final    router-ready

Columns are:

• persona id
• display name
• artifact status
• router readiness

/persona distill

Create a new persona from a corpus directory.

Example:

/persona distill mao_zedong data/MaoZeDongAnthology-master/src

What a successful distill means in this skill:

1. create the persona workspace
2. point it at the supplied corpus directory
3. read the corpus directly
4. select actual supporting sources
5. write a full persona.json
6. validate in final mode
7. build runtime modules
8. refresh the persona catalog
9. leave the persona router-ready

Important:

• success means final, not bootstrapped
• success means the artifact has evidence
• success means the artifact is usable for runtime switching
• the standalone CLI bootstrap helper is not the same thing as the full skill workflow

/persona delete

Delete a persona by id, display name, or alias.

Example:

/persona delete mao_zedong

If the target does not exist, the skill should say so explicitly.

/persona

Enter persona mode with the default runtime profile.

Example:

/persona mao_zedong

This is equivalent in intent to:

/persona switch mao_zedong

/persona

Enter persona mode with a scene hint so the runtime can choose a better profile.

Examples:

/persona mao_zedong analysis
/persona laozi default
/persona fozu conflict

Supported scene hints in the current runtime are typically:

• default
• decision
• conflict
• analysis

/persona switch [scene]

Explicit alias for persona activation.

Examples:

/persona switch mao_zedong
/persona switch mao_zedong analysis

/persona language

Show the current response language override.

Example:

/persona language

/persona language

Override the language used during persona replies without changing the persona’s style, method, or stance.

Examples:

/persona language English
/persona language 中文
/persona language 日本語

Meaning:

• the content language changes
• the persona style should remain the same
• if no override is active, roleplay uses the persona’s default language

/persona language default

Clear the active language override and return to the persona’s default language.

Examples:

/persona language default
/persona language reset

/persona fuse [name3]

Create a new fused final persona from 2-3 existing final personas.

Examples:

/persona fuse mao_hybrid mao_zedong laozi
/persona fuse strategic_mystic mao_zedong laozi fozu

Current fuse behavior:

• source personas must already be final
• target name must not already exist
• more than 3 sources is intentionally rejected by default
• fusion directly builds modules and runs final validation
• successful fusion ends in a router-ready persona

Important limitation:

• fusion is synthetic by nature
• it can produce a stable hybrid runtime artifact
• it should not be confused with a literal historical single person

/persona exit

Exit persona mode explicitly.

Equivalent exits:

/persona exit
/persona off
/persona quit
退出persona
退出角色
退出skill

Roleplay Behavior

When persona mode is active, the runtime contract expects:

• fully in-character replies
• first-person answers
• exact (name) prefix at the beginning of each in-character reply
• continued in-character behavior across turns
• no drift back to neutral assistant mode until explicit exit

If the user asks:

• 你是谁
• who are you

the runtime should answer as that persona.

If the user asks whether it is literally the real person or a simulation, the runtime should answer truthfully that it is a persona-based simulation, then continue according to the user’s requested mode.

Storage Layout

Main skill layout

• SKILL.md Public skill entrypoint for Codex.
• README.md This file.
• README_ZH.md Chinese version of the external README.
• scripts/cli/ Public command router for /persona ....
• scripts/extraction/ Scaffolding, validation, and runtime-module building for persona artifacts.
• scripts/runtime/ Catalog building, module selection, roleplay prompt rendering, and persona fusion.
• scripts/shared/ Shared helpers such as alias resolution and runtime state helpers.
• references/extraction/ Output contract, workflow, dimension catalog, and quality guidance for distillation.
• references/runtime/ Runtime and roleplay contracts.
• store/personas/ Persona artifact store.
• store/runtime/ Runtime session state, including language override session data.
• legacy/ Archived pre-merge extractor and roleplay folders kept only for traceability.

Persona store layout

Each persona usually lives under:

store/personas//

Typical contents:

• persona.json
• persona.template.json
• extractor_config.json
• modules/

Shared runtime index:

• store/personas/catalog.json

Runtime session file:

• store/runtime/session.json

End-to-End Examples

Example 1: Distill a new persona

/persona distill laozi data/LaoZiCorpus

Then verify:

/persona list

Then switch into it:

/persona laozi

Example 2: Switch, roleplay, exit

/persona mao_zedong
你的爱情观是怎样的
/persona exit

Expected runtime behavior:

• replies should start with (毛泽东)
• the persona should remain active until exit

Example 3: Use another language without changing persona style

/persona mao_zedong
/persona language English
What is your view of criticism?
/persona language default
/persona exit

Meaning:

• the answer should be in English
• the method and voice should still feel like the active persona

Example 4: Fuse two final personas

/persona fuse mao_lao_hybrid mao_zedong laozi
/persona list
/persona mao_lao_hybrid

Validation and Runtime Readiness

A persona should not be considered complete unless:

• artifact_meta.status == "final"
• source_index is populated
• evidence_index is populated
• dimensions contain real synthesis
• scenario_library is populated
• validate_persona.py --mode final passes
• module_registry exists
• load_profiles exists
• catalog marks it as router-ready

For this reason:

• bootstrapped is not complete
• draft is not complete
• empty evidence is not complete
• template shells are not complete

Important Behavior Differences

Distill

/persona distill at the skill level is intended to be end-to-end.

The standalone CLI helper:

python skills/persona/scripts/cli/persona_cli.py distill  

still behaves as the bootstrap stage used inside the skill workflow, not as the full agentic synthesis by itself.

Fuse

/persona fuse is now treated as a completion command.

That means it should:

• create the fused artifact
• build modules
• pass final validation
• refresh the catalog
• finish in router-ready state

Common Mistakes

• Treating legacy/ as an active second skill.
• Assuming bootstrapped means distillation is done.
• Expecting a persona with no evidence to roleplay reliably.
• Forgetting that language override should preserve persona style.
• Using fuse with non-final source personas.
• Using more than 3 source personas and expecting stable quality.
• Confusing scene position arguments with --scene flags.

Correct scene usage:

/persona mao_zedong analysis
/persona switch mao_zedong analysis

Current Active Paths

Default active Codex install path:

/Users/tomhu/.codex/skills/persona

Current workspace mirror in this environment:

/Users/tomhu/VsCode/test/skills/persona

If you move the skill to another machine, copy the whole skills/persona/ directory so the bundled store/ moves with it.

See Also

• SKILL.md
• references/extraction/output_contract.md
• references/extraction/agentic_distill_workflow.md
• references/runtime/roleplay_contract.md