persona-calibration（角色校准）

Persona Calibration

Use this skill to turn vague "I want a better personality" requests into a structured persona update workflow.

What This Skill Does

• Analyze common persona dimensions from industry examples before designing questions.
• Run a multi-round calibration interview instead of asking for a single freeform description.
• Separate persona core from expression style and operating rules.
• Convert interview results into concrete edits for SOUL.md, IDENTITY.md, USER.md, and optionally MEMORY.md.
• Prefer proposal → approval → edit for major personality changes.

When To Use

Use when the user asks to:

• create a new persona / personality for an OpenClaw instance
• refine or update an existing assistant personality
• make the agent more like a partner / advisor / operator / researcher / etc.
• compare persona frameworks or prompt dimensions from GitHub or the web
• design a questionnaire or interview for personality tuning
• help another person update their own OpenClaw instance persona

Core Principle

Do not jump straight into writing persona prose.

First extract the hidden dimensions behind the requested personality, then validate them through structured questions, then update files.

Workflow

Step 1: Baseline scan

If the user mentions existing persona repositories, prompt frameworks, or wants a more evidence-based process:

• Inspect a few representative sources from GitHub or the web.
• Extract recurring dimensions instead of copying wording.
• Summarize the dimensions in a compact model.

Read references/persona-dimensions.md for the default dimension model and example source patterns.

Step 2: Build a dimension model

Use these buckets as the default model:

• Identity / role
• Objective function / priorities
• Beliefs / philosophy
• Decision style
• Communication style
• Proactiveness boundaries
• Scenario behavior
• Anti-patterns / forbidden behaviors
• Memory policy
• Self-correction / update policy

Adjust if the user clearly needs more specific domains.

Step 3: Run calibration rounds

Do not dump a huge unstructured questionnaire all at once.

Use progressive rounds. The default order is:

• Round 1 — Core positioning

- role, priorities, judgment style, proactiveness

• Round 2 — Communication & expression

- length, structure, tone, disagreement style, forbidden phrasing

• Round 3 — Scenario calibration

- debugging, research, architecture, brainstorming, rushed tasks

• Round 4 — Boundaries, memory, correction

- what to remember, what not to remember, when to interrupt, how to update rules

• Round 5 — Example validation

- present several answer samples and ask for favorite / least favorite

Default size:

• 8-12 questions per round
• Mix direct preference questions with scenario questions and trade-off questions

Step 4: Ask better questions

Design questions to reveal not just stated preferences, but actual operating preferences.

Mix these question types:

• Baseline questions — explicit preferences
• Scenario questions — same dimension across multiple contexts
• Trade-off questions — force ranking between competing values
• Boundary questions — what feels annoying, wrong, or overbearing
• Example selection questions — choose preferred response samples

Important: do not hard-code a single number of clarification questions into the final persona. If the calibrated preference is to clarify first, phrase it as:

• ask 1-3 high-value clarification questions by default
• continue only if still needed
• stay concise; avoid turning the interaction into an interrogation

Step 5: Summarize after each round

After each round:

• give a concise interpretation of what the answers imply
• highlight any tension or unresolved ambiguity
• explain what the next round is trying to disambiguate

Do not silently absorb answers without reflecting them back.

Step 6: Produce a persona update proposal

Before editing files, produce a proposal that translates the interview into:

• role identity
• value hierarchy
• communication rules
• scenario handling rules
• memory rules
• update policy

For major personality edits, get explicit user approval before changing files.

Step 7: Apply updates to files

Map results to files like this:

• SOUL.md → personality core, communication style, behavioral rules, scenario guidance, absolute don'ts
• IDENTITY.md → concise role / creature / vibe summary
• USER.md → user preferences if the changes are really user-specific rather than agent-specific
• MEMORY.md → compact long-term summary of calibration results

Do not stuff every interview detail into every file. Keep long-term files crisp.

Step 8: Recommend a trial period

After editing:

• explain what changed
• recommend a short trial period in real conversations
• invite the user to note where the "feel" is still off
• suggest a second-pass refinement only after some real usage

Output Requirements

When reporting results, prefer this structure:

• Current read of the persona
• Key signals from the latest round
• Tensions / unresolved edges
• Proposed rule changes
• Next round or next action

When delivering the final proposal, include:

• final persona summary
• file mapping (SOUL.md, IDENTITY.md, USER.md, MEMORY.md)
• what should be edited now vs what should wait for trial feedback

Guardrails

• Do not optimize for theatrical prose; optimize for durable operating rules.
• Do not confuse personality with roleplay.
• Do not lock the persona into rigid behavior when the user actually wants conditional switching.
• Do not store low-value conversational fragments as long-term memory.
• Do not make major persona edits without a proposal review first, unless the user explicitly asked for direct changes.

References

• For default persona dimensions and what industry persona repos usually encode, read references/persona-dimensions.md.
• For a reusable round-by-round questionnaire scaffold, read references/questionnaire-template.md.
• For an example of how calibration results should be distributed across OpenClaw files, read references/file-mapping-example.md.