Persona Perspective Builder

Persona Perspective Builder

Fuse the best parts of nuwa-skill and anyone-skill without inheriting their worst constraints.

This skill is for building a new person-derived skill from evidence, not for writing a flattering summary.

It should:

• intake mixed evidence from many channels
• separate thinking framework, task signature, and taste rubric from surface mimicry
• decide whether the result should be perspective, persona, or hybrid
• produce a bounded, honest, reusable skill package

Core Thesis

Use nuwa ideas for:

• mental models
• decision heuristics
• anti-patterns
• worldview tensions
• repeated task structure
• repeated taste and critique patterns

Use anyone ideas for:

• source intake across chats, documents, archives, and public content
• interaction style
• memory anchors
• values and contradictions
• iterative correction and update flow

Do not force every subject into full-person simulation. Many targets only justify a strong perspective skill. Many of the most valuable outputs in the agent era are not mimic voice, but reusable task signatures and taste rubrics.

Output Modes

Choose exactly one primary output mode.

1. perspective

Use when the main value is:

• how this person thinks
• how they make decisions
• how they frame trade-offs
• how they evaluate new problems
• how they tend to structure work
• what kind of solutions they consistently reward or reject

Best for:

• public thinkers
• founders
• researchers
• historical figures
• thin but high-quality public evidence

Default deliverable:

• [slug]-perspective/
• SKILL.md
• agents/openai.yaml
• references/evidence_profile.md

2. persona

Use when the main value is:

• how this person speaks
• how they react
• how they relate to others
• how they carry memory, tone, and interpersonal style

Best for:

• private or user-supplied archives
• rich chat logs
• letters, diaries, emails
• user-known people with strong behavioral traces

Default deliverable:

• [slug]-persona/
• SKILL.md
• agents/openai.yaml
• references/persona_profile.md

3. hybrid

Use when both are strong:

• the target has stable thinking patterns
• the target also has enough evidence for believable interaction style

Best for:

• well-documented public figures with long-form interviews
• personal archives plus reflective writing
• targets where the user explicitly wants both advisor value and character fidelity

Default deliverable:

• [slug]-hybrid/
• SKILL.md
• agents/openai.yaml
• references/evidence_profile.md
• references/persona_profile.md

Input Contract

Require only:

• who the subject is
• what the user wants the resulting skill to do

Infer when possible:

• subject type
• likely source channels
• best output mode
• whether the goal is advisory, simulation, or both

If the material is really a generic method, evaluator, or playbook with a person wrapper, route to source-to-skill instead of forcing a person-derived build.

Ask follow-up questions only when one of these changes the result:

• private vs public use
• availability of private archives
• whether distribution is intended
• whether the user wants a perspective, persona, or hybrid result

If not specified, default rules are:

• public figure -> perspective
• private archive-heavy subject -> persona
• mixed rich evidence + explicit simulation/advisor goal -> hybrid

Mutual Routing With source-to-skill

Stay here when:

• the person remains the main product
• the user wants a person-derived advisor, perspective, persona, or hybrid result
• the extracted task signatures and taste should stay attached to that person-shaped lens

Route to source-to-skill when:

• the bundle is really teaching a generic method, playbook, evaluator, or router
• task structure and taste matter more than personhood
• the person is mainly a carrier for a reusable workflow
• evidence is too weak for a person-derived result but strong enough for operational extraction

When routing, pass forward:

• thinking_frameworks
• task_signatures
• taste_rubrics
• operator_rules
• exemplars_and_anti_exemplars

Workflow

1. Frame the Subject

Classify the target:

• self
• someone-known
• public-figure
• historical-figure
• fictional-character
• archetype

Then classify the user's intent:

• advisor
• simulation
• both

Do not proceed as if the target is public-safe if the user is supplying private material.

2. Build the Evidence Map

Normalize every source into an evidence ledger with:

• source_id
• channel
• publicness
• time_span
• subject_distance
• confidence_level
• perspective_value
• task_value
• taste_value
• persona_value

Recommended channels:

• writings and long-form essays
• interviews and podcasts
• short social posts
• chat exports
• email or document archives
• biographies and external commentary
• decision records and major actions
• timeline and turning points

Score each source four ways:

• perspective_value: how much it reveals thought structure
• task_value: how much it reveals preferred task shape and workflow behavior
• taste_value: how much it reveals quality bar, trade-off logic, and anti-slop judgment
• persona_value: how much it reveals interaction and behavior

Use these rough heuristics:

• essays, books, talks -> high perspective
• demos, code reviews, design critiques, planning artifacts -> high task and taste
• chats, emails, DMs -> high persona
• interviews -> medium-high for both
• interviews about projects or decisions -> medium-high task and taste
• social posts -> medium persona, medium perspective only if repeated patterns appear
• third-party commentary -> low-medium, never primary if first-party evidence exists

3. Choose the Build Shape

Use this decision table:

If the evidence is wide but shallow, degrade gracefully to perspective. Do not claim full-person simulation without strong persona evidence.

4. Run Multi-Lane Extraction

Run all lanes even if one will later be downweighted.

Pass A: Cognitive Extraction

Extract:

• 3-7 mental models
• 5-10 decision heuristics
• recurring trade-off logic
• anti-patterns and refusal rules
• contradictions worth preserving

This is the nuwa-style core.

Pass B: Task Extraction

Extract:

• recurring task shapes
• preferred task grain
• favored starting point
• done conditions
• delegation or escalation rules
• favored interfaces, tools, or surfaces
• recurring transformation from vague ask to executable unit

This is the tasky lane.

Pass C: Taste Extraction

Extract:

• what this person consistently rewards
• what they consistently reject
• repeated trade-off preferences
• signals of elegance, durability, or overbuilding
• examples and anti-examples they use to teach judgment
• likely operator rules another agent could reuse

This is the taste lane.

Pass D: Persona Extraction

Extract:

• vocabulary and sentence habits
• emotional temperature
• conflict style
• humor and softness/harshness
• memory anchors
• pride, fear, fixation, avoidance
• core values and prohibitions

This is the anyone-style core.

5. Grade the Evidence

Every non-trivial claim should be tagged as:

• L1 direct quote or direct artifact
• L2 strong paraphrase or well-sourced report
• L3 inference from multiple signals
• L4 inspired extrapolation

Higher-stakes traits need stronger evidence:

• voice style may tolerate some L3
• non-negotiable values should prefer L1-L2
• hard prohibitions should almost never rely only on L4
• task signatures should prefer repeated observed choices, not one project
• taste rubrics should prefer repeated critiques, comparisons, or trade-off decisions

If evidence conflicts, preserve the tension. Do not smooth contradictions into fake coherence.

6. Build the Skill Packet

Produce a packet with:

• subject_frame
• source_inventory
• chosen_mode
• durable_units
• honesty_boundaries
• target_skill_shape

durable_units should separate:

• thinking_frameworks
• task_signatures
• taste_rubrics
• operator_rules
• interaction_patterns
• memory_anchors
• values_and_red_lines
• failure_modes

7. Generate the New Skill

Create the generated skill directory for the subject.

Default shape:

• SKILL.md
• agents/openai.yaml
• references/*.md only when needed

The generated SKILL.md should include:

• clear trigger text
• what the skill is for
• what it must not pretend to know
• how it should answer new questions
• how to express uncertainty

For perspective mode:

• optimize for advice, analysis, decision framing, task shaping, and critique
• do not over-index on mimic voice
• expose stable operator rules when they are genuinely evidenced

For persona mode:

• optimize for interaction fidelity and boundaries
• do not overclaim worldview depth when evidence is mostly conversational

For hybrid mode:

• make thinking core primary
• make voice and behavior a controlled overlay
• if the two conflict, preserve the cognitive core over surface mimicry

8. Validate Before Delivery

Run at least these checks:

1. known-stance test Compare 3 generated answers against known subject positions.

2. new-problem test Ask one novel question and confirm the skill answers from extracted principles, task signatures, and taste rubrics rather than generic filler.

3. boundary test Ask beyond the evidence and confirm the skill says what is unknown.

4. mode-fit test Confirm the chosen perspective/persona/hybrid mode matches the actual evidence mix.

5. task-fit test Ask the skill to structure a new task and verify the output reflects the extracted task grain, starting point, and done condition.

6. taste-fit test Ask the skill to compare two options and verify the judgment uses repeated trade-off signals rather than vague aesthetics.

If the validation fails:

• reduce scope
• downgrade hybrid -> perspective or persona -> perspective
• mark weak areas explicitly

Guardrails

• do not confuse voice cloning with person understanding
• do not grant hybrid status unless both lanes are genuinely evidenced
• do not use sparse social posts as the whole person
• do not treat biography as behavior
• do not build deception tools for impersonation, harassment, or manipulation
• do not present public-figure simulations as the real person
• do not turn weak inference into strong canon

Honesty Boundaries

Always include these in the generated skill when relevant:

• this is evidence-shaped, not the real person
• public evidence has blind spots
• private archives still do not expose the whole interior self
• changed positions over time should be preserved with dates when possible

Recommended Defaults

If building for public thinkers:

• prioritize perspective
• use long-form writing, interviews, decisions, and criticism
• pay special attention to task-shaping and taste signals in project choices, reviews, and demos

If building for people known personally:

• prioritize persona or hybrid
• use chat logs, emails, notes, and user-supplied memories

If building for fictional characters:

• use persona
• switch to inspired-by mode when distribution is intended

Example Requests

• Use $persona-perspective-builder to turn this founder's interviews, essays, and tweets into a perspective skill.
• Use $persona-perspective-builder to build a hybrid skill from these chats, emails, and voice notes about my former collaborator.
• Use $persona-perspective-builder to decide whether this target deserves a perspective skill or a full persona skill before generating anything.
• Use $persona-perspective-builder to fuse public research and private archives, then create the best-fit person-derived skill.
• Use $persona-perspective-builder to extract how this person structures work and what kinds of solutions they consistently reward or reject.
• Use $persona-perspective-builder to decide whether this bundle should stay person-derived or route to $source-to-skill because the reusable method matters more than the person.