persona

name	persona
description	Unified persona skill for distilling, listing, fusing, switching, and exiting personas. Use this skill for all persona workflows through `/persona ...`.
argument-hint	<subcommand-or-persona-name> ...
version	2.0.0
user-invocable	true

Persona

This is the single public entrypoint for all persona workflows.

Use /persona ... for:

• help
• language
• distill
• delete
• list
• fuse
• switch
• exit

Primary runtime data:

• store/personas/catalog.json

Primary scripts:

• scripts/cli/persona_cli.py
• scripts/extraction/scaffold_persona_project.py
• scripts/extraction/validate_persona.py
• scripts/extraction/build_persona_modules.py
• scripts/runtime/build_persona_catalog.py
• scripts/runtime/select_persona_modules.py

Read these references when needed:

• references/extraction/output_contract.md
• references/runtime/roleplay_contract.md
• references/extraction/dimension_catalog.md
• references/extraction/agentic_distill_workflow.md
• references/extraction/distill_execution_prompt.md
• references/extraction/pipeline.md
• references/extraction/quality_rubric.md

Commands

1. Help

Show all available persona commands:

/persona help

Use:

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

2. Language

Show or set the active response language:

/persona language
/persona language 中文
/persona language English
/persona language default

Use:

python skills/persona/scripts/cli/persona_cli.py language
python skills/persona/scripts/cli/persona_cli.py language English
python skills/persona/scripts/cli/persona_cli.py language default

If no override is active, roleplay uses the persona default language.

3. Distill

Create or refresh a persona workspace from a supplied corpus directory:

/persona distill <persona_name> <data_dir>

Example:

/persona distill mao_zedong data/MaoZeDongAnthology-master/src

What this should do:

1. create store/personas/<persona_name>/
2. scaffold the project
3. set corpus_root
4. inspect the raw corpus directory directly
5. choose source-grounded dimensions, evidence, voice, and scenarios
6. write or rewrite persona.json
7. validate the artifact
8. build runtime modules
9. refresh the catalog
10. finish only when the persona is usable or a concrete blocker has been identified

Use:

python skills/persona/scripts/cli/persona_cli.py distill <persona_name> <data_dir>

Important:

• /persona distill is an end-to-end skill workflow
• do not stop after bootstrap unless the user explicitly asked for initialization only
• the CLI script above is the bootstrap helper used inside the workflow, not the whole workflow by itself
• the scaffold should provide a rich multi-dimensional template, not an empty shell
• the agent should prune unsupported dimensions and fill supported ones from direct corpus reading
• it does not generate candidate passages, block indexes, keyword search packs, or heuristic drafts
• a bootstrapped or draft persona is not a successful /persona distill result unless the user explicitly asked to stop early

If a persona with the same name already exists:

• do not overwrite silently
• in the interactive skill workflow, ask the user whether to delete and rebuild
• if the operator is using the CLI helper directly, fail with a clear message and point them to /persona delete <name>

Distill Rule

For /persona distill, the extraction method is agentic direct reading.

Before doing any work, load and obey:

• references/extraction/distill_execution_prompt.md

The assistant should:

1. bootstrap the workspace with the CLI helper if the persona directory does not already exist or needs rebuilding
2. if a persona with the same name already exists, ask the user whether to delete and rebuild instead of overwriting silently
3. read the raw corpus directory directly
4. choose dimensions based on the actual source material instead of forcing a fixed template
5. manually author source_index and evidence_index only for the sources and excerpts actually used
6. write or rewrite persona.json from direct source understanding
7. run final validation
8. build runtime modules
9. refresh the catalog
10. confirm whether the persona is now router-ready

Unless the user explicitly asks to stop early, /persona distill should complete the full extraction in one command flow.

For actual execution, this means:

• after bootstrap, immediately inspect the corpus directory and start reading source files in the same turn
• do not wait for another user message such as “继续蒸馏”
• do not summarize a plan and stop
• do not hand the user a partially prepared workspace as if that fulfilled the command
• keep writing and validating until the artifact reaches final and becomes router-ready

Distill Acceptance Criteria

Treat /persona distill as incomplete unless all of the following are true:

1. persona.json is populated from direct corpus reading rather than left as a template shell
2. source_index is non-empty and reflects the actual source files used
3. evidence_index is non-empty and each important dimension is anchored to real evidence
4. there are enough filled dimensions to model the persona in detail, not just a few generic traits
5. scenario_library is populated with representative source-grounded or clearly marked synthetic cases
6. validate_persona.py --mode final passes
7. build_persona_modules.py has been run
8. the catalog has been refreshed
9. the persona is router-ready

Do not answer as if distillation succeeded when the artifact is only a scaffold, only a bootstrap, or missing evidence.

Do not generate or rely on candidate_passages.json, candidate_evidence.md, block_index.json, or heuristic draft synthesis.

Distill Enforcement

When the user invokes /persona distill <name> <data_dir>, treat it as an execution order, not a planning request.

Required behavior:

1. do not stop after calling the bootstrap CLI helper
2. do not return a “next steps” message as the main result
3. do not claim success until final validation passes and the persona is router-ready
4. continue reading files, writing persona.json, validating, and rebuilding until either:
   • the distill completes successfully, or
   • a concrete blocker is found that cannot be resolved within the current turn
5. if a blocker occurs, report the blocker precisely and include the exact file, command, or validation failure

Forbidden completion states for /persona distill:

• bootstrapped
• draft without passing final validation
• empty source_index
• empty evidence_index
• empty or near-empty scenario_library
• template-like dimensions with blank synthesis summaries

Forbidden assistant outcomes for /persona distill:

• replying mainly with “bootstrap ready”
• replying mainly with “next steps”
• asking the user to manually continue distillation in a follow-up turn
• claiming the command succeeded when the persona is not final
• switching the persona before final validation passes

4. Delete

Delete a distilled persona:

/persona delete <persona_name>

Use:

python skills/persona/scripts/cli/persona_cli.py delete <persona_name>

Rules:

• if the persona does not exist, say it does not exist
• do not delete the built-in starter_template

5. List

List all distilled personas:

/persona list

By default this should list distilled personas and hide template personas.

Use:

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

6. Switch / Use

Enter persona mode:

/persona <persona_name>
/persona <persona_name> <scene>
/persona switch <persona_name>
/persona switch <persona_name> <scene>

Supported scene hints:

• default
• decision
• conflict
• analysis

Example:

/persona mao_zedong
/persona mao_zedong conflict

Use:

python skills/persona/scripts/cli/persona_cli.py use <persona_name> [scene]

Explicit switch alias:

python skills/persona/scripts/cli/persona_cli.py switch <persona_name> [scene]

7. Exit

Leave persona mode:

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

8. Fuse

Create a fused final persona from 2-3 existing final personas:

/persona fuse <new_persona_name> <name1> <name2>
/persona fuse <new_persona_name> <name1> <name2> <name3>

Example:

/persona fuse mao_hybrid mao_zedong lu_xun

Use:

python skills/persona/scripts/cli/persona_cli.py fuse <new_name> <name1> <name2> [name3]

More than 3 source personas is not supported by default. If the user insists, warn that output quality is not guaranteed.

Rules:

• do not overwrite an existing target persona silently
• source personas should already be final
• /persona fuse should finish by building modules, passing final validation, refreshing the catalog, and making the fused persona router-ready
• if the fused artifact cannot pass final validation, report the exact blocker instead of calling it complete

Router Behavior

If the first argument matches:

• help
• language
• list
• distill
• delete
• fuse
• exit
• off
• quit

then treat it as a subcommand.

Otherwise treat it as a persona switch request:

/persona <persona_name> [scene]

Roleplay Rule

When persona mode is active:

• remain fully in character until exit
• answer in first person as the persona in ordinary dialogue
• if asked 你是谁 or who are you, answer as the persona would identify themselves
• if asked whether you are literally the real person, the underlying model, or a simulation, answer truthfully that this is a persona-based simulation, then continue according to the user's requested mode
• do not drift back to neutral assistant style unless the user exits or explicitly requests out-of-character analysis
• if /persona language <language> is active, answer in that language while preserving persona voice
• if no override is active, answer in the persona default language

Runtime Selection

For switching into a persona:

1. resolve persona from store/personas/catalog.json
2. choose a load profile by explicit scene or scene heuristic
3. load only the modules required for that profile
4. avoid loading the full artifact unless deep analysis is required

Agentic Distill

When extracting a new persona, prefer this order of authority:

1. direct reading of source texts in the user-provided corpus directory
2. the scaffold dimension template as a modeling checklist
3. validation failures as feedback on what still needs to be authored

The extraction workflow should be controlled by the assistant's reading and judgment, not by fixed keyword rules alone.

Internal Layout

This skill is organized as one unified directory:

• scripts/cli/ for public command routing
• scripts/extraction/ for distillation and validation
• scripts/runtime/ for switching, cataloging, prompt rendering, and fuse
• scripts/shared/ for shared path resolution helpers
• store/personas/ for portable distilled persona artifacts bundled with the skill
• references/extraction/ for extraction contracts and modeling references
• references/runtime/ for roleplay/runtime rules
• legacy/ for archived pre-merge skill folders retained only for reference