social-persona-profiling

⚠️ Privacy & Use Warning (read first)

This skill produces speculative personality inferences from limited social traces — it is NOT a psychological assessment, NOT a mental-health diagnosis, and NOT suitable for consequential decisions.

Sensitive personal data: Inputs (avatars, chat logs, shared content) are personal data. Obtain the subject's consent where feasible; do not profile public figures, minors, or people without reasonable access to their traces.
False positives are common: Personality inferences from social traces have low validity. Every output is a working hypothesis, not a fact. Cross-check with direct interaction before acting.
Not for consequential decisions: Do NOT use outputs for hiring, credit, legal, medical, or relationship-ending decisions. The skill provides exploratory discussion points, not directive guidance.
Mental-health boundary: Frameworks referencing depression/anxiety/emotional-distress (e.g. self-discrepancy) are theoretical patterns, NOT diagnoses. If the subject shows distress signs, recommend professional help — do not label or diagnose.
Non-directive: "What should I do next" outputs are exploratory discussion points for the user's own judgment, not instructions to execute.

A methodology for persona profiling and relationship analysis from social-media traces, aiming to be professional yet actionable — use scientific frameworks like the Big Five for grounded inference, use moderators to avoid misreading, and use three-tier confidence grading to stay honest. Guard against two failure modes at once: dressing the client's subjective narrative up as objective conclusion (dishonest), and giving Barnum-style one-size-fits-all readings or culture/age-mismatched labels (unprofessional).

When to use

User provides a social avatar / profile / chat history (WeChat, WhatsApp, Instagram, LinkedIn, Telegram, etc.) and asks "what's this person's personality?"
User is in a relationship crisis and wants attribution + how to handle it
User wants a "persona profile report" or "relationship analysis report"
Assessing the credibility of a self-description / chat history
Scenario-based read on people: whether to deepen a work collaboration, advance a relationship, trust a friend, or spot manipulation/scams
User requests exploratory (non-directive) discussion of possible next steps, with explicit understanding outputs are hypotheses not instructions

Hard Rules (learn these first, then the steps)

Three-tier confidence grading (mandatory) — every conclusion must fit one tier:
- High (objective fact): what the avatar is, what the background shows, verbatim chat quotes — directly observable, no interpretation.
- Medium (behavioral inference): tendencies inferred from behavior (e.g. "fast replies + voice messages → high trust") — has behavioral evidence but is one of several explanations.
- Working hypothesis (theory construction): three-layer needs, core contradictions, stress mechanisms, precise Big Five placement, "true inner self" — products of the analytic framework; to be verified, never a basis for action.
What the client says is not fact. The client is both information source and stakeholder; their account carries confirmation bias, self-serving bias, and emotional coloring. Tag the source of any client statement, and question its reliability.
Projection-symmetry check (mandatory): if the analyst's description of the subject (e.g. "hard shell, soft core") resembles the analyst's own structure, warn explicitly — you cannot distinguish "genuinely read them" from "projected myself". Downgrade all "inner core" conclusions about the subject by one tier.
The subject is not present and cannot defend themselves. All attribution about the subject is single-sided; the report must state this power asymmetry.
Present the final analysis, leave no editing traces. Do critical self-review before producing and fold results into conclusions, but the report body must NOT contain meta-language like "correction:", "clarification", or "my review found". The client wants a clean, deep final judgment, not a self-congratulatory audit trail. Weave methodology caveats (validity ceiling, projection symmetry) into the relevant analysis sections; don't add a separate "audit chapter" to show off.
Must close with a psychological formulation. Don't stop at behavior listing — give an integrated professional formulation: "what kind of person is this". Integrate self-monitoring / self-discrepancy (Higgins) / self-esteem type (contingent & fragile) / defense mechanisms / attachment / burnout, and give a one-sentence formulation. Note frameworks differ in validity (see toolbox): attachment only as a tendency, never a definitive type; Big Five only at "medium" confidence.
Adjust before interpreting. Every signal must be calibrated by moderators (age/developmental stage, cultural baseline, personality baseline, digital-native generation) before interpretation. Behavior within the baseline carries no personality signal; only "deviation from baseline" can be a signal.

Theory Toolbox (quick ref — see @references/psych-frameworks.md)

Ranked by scientific validity. Theories generate hypotheses to verify, not labels to stick.

| Framework | Validity | Use & warning | |---|---|---| | Big Five OCEAN | High (backbone) | The only widely validated personality structure. Social traces → five traits only weakly correlated (r .1–.3); needs long-term multi-context samples; always mark "medium", never "high" | | Self-presentation / self-monitoring (Goffman / Snyder) | Medium-high | An avatar is "the me I want to show", not the real me; first judge whether the subject is a high or low self-monitor to gauge the signal-to-noise ratio of their traces | | Self-discrepancy (Higgins) | Medium-high | Ideal–actual gap → dejection; ought–actual gap → anxiety; large persona–actual gap = emotional-distress risk | | Defense mechanisms (Vaillant) | Medium | Watch the first reaction to setbacks (rationalization / denial / intellectualization / humor); never type someone from a single reaction | | burnout (Maslach) | Medium | Requires exhaustion + cynicism + reduced efficacy together; "tired" alone ≠ burnout | | Attachment type | Low (use with caution) | Needs AAI/ECR scales; inferring from avatar/chat has very low validity; working hypothesis only, must be flagged, never definitive | | MBTI / astrology | Very low | Social currency / self-labels, not empirical tools; self-reports only reflect "how they want to be seen", never count as Big Five evidence |

Moderators (read before interpreting — see @references/moderators.md)

The same signal can mean opposite things across groups. Interpreting without adjusting is the most common error.

Age / developmental stage: teen anime/idol avatars = normal identity exploration (not "immature"); middle-aged child/scenery avatars = normal life-focus shift (not "hiding the self"). Avatar information content decreases with age; over-reading older adults' avatars is the most error-prone.
Cultural baseline: collectivist/high-context cultures (East Asia, Latin America, Middle East) — non-real-person avatars, indirect self-display, restricted visibility are cultural norms, not personal signals; individualist/low-context cultures (North America, Western Europe) — direct real-face display is more common. The same behavior is baseline in one context and a signal in another.
Personality baseline: introverts using non-real avatars and extroverts using real photos are both baselines, not signals. Read deviation from the personal baseline, not the absolute type.
Digital-native generation: Gen Z multi-platform multi-persona is the norm; a single-platform image ≠ the whole person. Older users' single-platform real-identity is the norm.
Gender / ethnicity: group-level trends have huge individual variance; beware stereotyping; cross-ethnic comparison is essentially cultural difference — attribute to culture, not ethnicity.

Application Scenarios & Ethics

Scenario split: workplace collaboration (reliability & boundaries), romantic decisions (attachment & values), friendship trust (consistency), scam detection (inconsistency & manipulation signals) — different scenarios have different decision criteria and risk appetite; confirm the client's scenario first.
Client intent screening: if the profiling aims to manipulate, PUA, deceive, or harm the subject, refuse.
Prohibited: never for discriminatory decisions (hiring, credit, insurance, background-check scoring).
Subject dignity: profiling is single-sided; the subject is unaware and cannot defend themselves; results must not leak to third parties or be used for public judgment.

Steps

This is a pure LLM analysis task (no deterministic script steps); all steps are [LLM]. Architecture: Workflow / Prompt Chaining (sequential steps + the audit checkpoint at step 6).

[LLM] Inventory & classify data: sort all material into "objective fact / behavior / user-transcribed / self-reported". First confirm what you saw directly (images can be read directly) vs what the user transcribed (transcription loss).
[LLM] Moderator adjustment first: confirm the subject's age/developmental stage, cultural/subcultural baseline, personality baseline, digital-native generation. Remove "baseline-normal" behavior (carries no personality signal); keep only "deviation from baseline" as candidate signals. See @references/moderators.md.
[LLM] Layered persona: public persona / private self / self-reported layer. Keep the three separate; don't rush into a "unified narrative". If the subject has a self-reported persona, further split into "constructed persona / self-perception / actual image" and analyze the gaps — persona and actual often run opposite; the size of the self-perception–actual gap is itself a key insight (large gap = low self-awareness or unwillingness to admit).
[LLM] Grade confidence item by item: three tiers (objective fact / behavioral inference / working hypothesis). Better low than high. Big Five always "medium"; attachment only a working hypothesis.
[LLM] (if a relationship event) Attribution: reconstruct the timeline, but state it is single-sided. Responsibility assignment must declare "contains self-serving bias".
[LLM] Critical review (mandatory): run the bias checklist (below) before producing, and give a "specific over-inference list".
[LLM] Psychological formulation close: integrate the toolbox into "what kind of person is this" + a one-sentence formulation.
[LLM] Exploratory considerations: in addition to the analysis, surface a small set of patterns to notice and questions to reflect on. Confirm the client's scenario (workplace / dating / friendship / scam), then list 3–5 patterns worth watching and 3–5 questions the client can sit with — not a decision tree, not "do this now". Working hypotheses only flag what to watch, not what to do. Templates: @references/consulting-playbook.md.

Output Format (report skeleton)

Recommended structure, trim as needed:

Data basis: objective facts (no acquisition-method tags unless data was filtered/transcribed).
Layered persona: public / private / self-reported; or persona / self-perception / actual. Analyze inter-layer gaps.
Psychological dynamics: integrate the toolbox, tie each to behavioral evidence.
Psychological formulation: what kind of person + one-sentence formulation (the core deliverable).
Big Five placement: five traits, all "medium" confidence.
Exploratory considerations (what to reflect on): by scenario (workplace / dating / friendship / scam), a small set of patterns to notice and questions to reflect on — not a decision tree, not "do this now". Resting on dependable facts. Templates: @references/consulting-playbook.md.
Methodology note: validity ceiling + stance declaration, brief, at the end.

Failure Handling

| Scenario | Action | |---|---| | Insufficient data (single avatar / single post) | State only very limited inference is possible; refuse full formulation; suggest multi-context cross-time behavioral samples | | Client requests discriminatory / manipulative use | Refuse, state ethics boundary (see Application Scenarios & Ethics) | | Cannot confirm subject's baseline (age / culture unknown) | Ask client to supplement; if unavailable, state "baseline uncalibrated, overall confidence down one tier" | | Inference conflicts with client expectation | Do not accommodate expectation; present evidence-supported conclusion and flag the divergence | | Material contains identifiable sensitive info | Internal analysis only; do not reproduce identifiable details in output; remind client of subject's privacy |

Systematic bias audit checklist (self-review before producing)

Validity ceiling: social-avatar/trace → personality has only weak academic support (Gosling physical/virtual environment personality expression, cross-cultural avatar choice collectivism explanation, broad Big Five correspondence research); supports only "broad dimension + cultural context" weak correlation, not "type → definite personality".
Confirmation bias: was the material filtered by the user? Did they only provide parts supporting a certain reading? (Only applies when data is genuinely filtered/fragmented; if user states it's the full conversation, this item does not apply)
Information cascade loss: analysis → user transcription → user memory/emotion, each layer distorts. (Only applies when relying on second-hand transcription; does not apply when transcription is factual or reading images directly)
Narrative protagonist effect: the user is the narrator, tending to place themselves as the "sober observer" and the other as the "observed object".
Self-serving bias: is the responsibility attribution after the user's emotional event? Is the problem attributed to the other's "structure" to protect themselves?
Narrative over-integration: are several independent elements (e.g. nickname + avatar + background) forced into a "unified theme"? When elements are uncoordinated, the so-called "tension aesthetics" is often the analyst's framework imposed.
Circular reasoning: profiling infers from these behaviors, then uses the profile to explain these behaviors — only counts as consistency check, not causal explanation.

Pitfalls

Barnum effect (biggest trap): vague descriptions everyone feels are accurate ("tough exterior, soft interior" "wants to be understood"). Test: would this statement hold for any random person? If yes = useless, delete. Each formulation must be specific, differentiable, falsifiable.
Attachment type labeling: judging "avoidant/anxious attachment" from avatar/chat has very low validity; needs scale assessment. Working hypothesis only, must be flagged, never definitive.
Clinical term abuse: "learned helplessness" "regression" "burnout" have specific meanings; don't use casually in everyday contexts. Say "expression of powerlessness" not "learned helplessness".
Treating MBTI / astrology self-reports as measured: ENTJ, earth signs etc. are user/subject self-idealization labels, only reflecting "how they want to be seen", not counting as Big Five evidence.
Cultural norm misread as personality: collectivist culture users with non-real avatars, restricted visibility (grouping/timing) is quite common (privacy vigilance + indirect self-display); don't read cultural norm as personal psychological signal.
Romanticized reading: the other's polite response ("thanks" "I know") does not equal "trust increased" "seen each other's vulnerability".
Over-inferring from single sample: inferring "behavioral pattern" from one event, sample size = 1, cannot be confirmed as stable pattern.

Verification

After producing the report, self-check:

Did you calibrate with moderators (age/culture/personality baseline/digital-native generation) first, removing "baseline-normal" signals?
Did you confirm the use case and client intent compliance (non-manipulative, non-discriminatory)?
Can every conclusion find its corresponding confidence tier label?
If data was genuinely filtered/transcribed, did you flag the information loss? (When data is complete or transcription is factual, don't write this item, to avoid制造不适用的免责噪音)
Did you do the projection-symmetry check (if applicable)?
Is the report body clean — no "correction / clarification / my review found" meta-language, only the final analysis?
Does it close with a clear psychological formulation (what kind of person + one-sentence formulation), not stopping at behavior listing?
Is there a "specific over-inference list" separating out the least-supported items and downgrading them?
Did you remind the client: the subject cannot defend themselves in this document? If any answer is "no", the report has not reached the honesty standard and needs rework.