关系沙盘｜十六型人格的互动流转 Skill

概述

关系沙盘｜十六型人格互动推演 v1.1.9

你可以直接这样说

• 把我和伴侣的 MBTI 关系做成一个互动图。
• 我们团队有 INFP、ENTJ、ISTJ，模拟一下怎么协作。
• 为什么我们总在沟通上误解？用 MBTI 给几个可能路径。

生成一个可编辑的自包含 HTML 关系沙盘，用来观察小型或多人 MBTI 关系网络中的可能互动。交付应是可操作的关系推演页面，不是静态文字报告；任何 Agent 都应能按说明生成或复用页面，不依赖 Codex 专用接口。

必须保留的边界说明

生成页面前或页面内必须明确写出：

• MBTI 只作为叙事和自我观察框架，不是科学诊断或确定性预测系统。
• 沙盘输出的是可能路径，不是事实、命运或专业心理建议。
• 避免“必然”“一定会”“这种类型总是如此”等绝对表述。
• If the user mentions abuse, coercion, self-harm, severe psychological distress, legal decisions, medical decisions, or major financial decisions, stop MBTI forecasting and provide safety-oriented guidance or recommend qualified professional help.

Use probabilistic language: "may", "tends to", "under stress may", "one plausible branch is".

Quick Start

1. Gather or infer these inputs:
   • People nodes: 2-20 people, each with name/label, MBTI type, and optional personality notes.
   • Relationship links: user-defined links from one person to one or more other people, with current relationship, desired relationship, and concrete event notes. Do not force sequential A-B-C relationships.
   • Expected relationship target: optional target group such as A-C or A-C-H, expected relationship, and a confidence question.
   • Relationship context: romance, ambiguous relationship, friendship, family, coworker, founder partnership, roommate, client relationship, competition, online, secret, or custom.
   • Current state: early attraction, stable intimacy, boundary negotiation, low-frequency contact, conflict freeze, repair window, or unknown.
   • Focus question: future direction, conflict risk, repair strategy, random events, long-term fit, communication style, or custom.
2. Generate the HTML sandbox with the bundled generator:

python3 scripts/generate_sandbox.py --output-dir ./sandbox-out

The script writes relationship-dynamics-sandbox.html into the output directory and prints its path. To inject starter inputs, pass a JSON object, for example --seed-json '{"people":[{"name":"A","type":"INFJ"}]}'. MBTI values are restricted to the 16 standard four-letter types; invalid seed types fail before a page is written.

1. Open or link the generated HTML file for the user.
2. Summarize what the page can do: advance phases, trigger weighted random events, trigger black-swan events, watch state transitions, and inspect the event ledger.

If the user provides incomplete inputs, create sensible defaults and keep the UI editable.

Dynamic Sandbox Requirements

The generated interface must include:

• A people-node editor for 2-20 MBTI nodes. Default to exactly 2 people. Let users add MBTI nodes one by one up to 20, and allow deleting added nodes while keeping the 2-person minimum.
• A relationship-link dialog that opens from the people-node area even with the default two people. It should let users choose a source person, connect to one or more target people, set current relationship, set desired relationship, and add concrete event notes.
• An expected-relationship validation system. Users can ask whether A-C, A-C-H, or any selected group can reach an expected relationship. Output a clear yes/no/conditional answer with confidence and plain-language reasons.
• A relationship state field using professional terms, not casual "first-line/second-line" labels.
• A visual simulation zone showing distance, tension, trust, attachment, volatility, repair capacity, misunderstanding load, and emotional charge.
• Weighted random events that depend on MBTI tendencies, relationship context, current state, user-defined relationship links, expected relationship target, and accumulated history.
• Low-probability black-swan events that can sharply change variables.
• Event cards during simulation. Each reasoning point must show at least 3 concrete events and may show 4-6 events. Events should appear dynamically, generally one by one with about a 2-second stagger, and remain inspectable. If many events appear, later events may appear in small parallel batches.
• A visible pause/continue control during simulation. Users must be able to pause, inspect event cards, click an event, edit it, and continue the simulation. Edited events must feed into later reasoning, not only change display text.
• Event cards should appear in randomized non-overlapping positions as much as practical.
• An event ledger explaining what happened, who was involved, which concrete case occurred, and why it affected the relationship.
• A future branch panel showing likely next states and suggested repair or escalation routes.

Conclusion Writing Rules

Conclusions must be beginner-friendly, concrete, and subject-specific:

• Every concrete case must name the participants and MBTI types, for example "A（INTP）和 C（INTJ）".
• Do not write generic claims without a concrete case.
• Start with a direct relationship forecast: who and who will likely move toward what direction, because of which concrete variable or event.
• Treat the user's relationship goal (for example exit, repair, cooperate, slow down, set boundaries) as the main conclusion axis. Do not confuse it with optional expected-relationship validation.
• Decompose user-written descriptions into causal variables such as responsibility load, resource value, anxiety, boundary risk, hidden actions, reliability, repair behavior, and external pressure. Do not merely quote the user's text.
• Remove filler such as "this is not fortune telling", "I looked at these people", "I will not guess", or generic disclaimers from result cards. Keep only words that change the user's decision.
• Keep examples, placeholders, and UI hints out of inference input. Empty fields must remain empty variables.
• Start from life-like examples, then summarize in plain language.
• Avoid overly technical language unless it is inside a clearly separated "logic/audit" section.
• If a conclusion category has no useful evidence, omit it or state that evidence is insufficient.
• Group the result page into colored sections such as target validation, whole-network relationship, multi-angle judgment, event chain, and reasoning audit.

Universal Agent Compatibility

This skill should be portable across Agent platforms:

• Prefer a single self-contained HTML file with embedded CSS and JavaScript.
• Do not require proprietary Agent tools. If browser automation is unavailable, provide the generated file path and concise usage instructions.
• If the platform cannot run local scripts, copy assets/sandbox.html as the deliverable directly.
• Keep data and inference rules in plain JavaScript objects so another Agent can inspect or modify them.
• Do not depend on hidden state from a previous Agent run. The page must be editable by the user.

Relationship State Terms

Use these states unless the user asks for a different taxonomy:

• High-Intimacy Bond: high access, high emotional salience, strong expectations.
• Stable Alliance: reliable connection with moderate boundaries and lower emotional volatility.
• Boundary Negotiation: active recalibration of closeness, expectations, roles, or power.
• Low-Frequency Link: contact remains possible, but emotional investment and access are reduced.
• Conflict Freeze: unresolved rupture, avoidance, punishment, shutdown, or cold stalemate.
• Repair Window: a temporary opening where clarification, apology, or shared action may restore trust.
• Exit Drift: gradual disengagement with low repair energy.

Randomness Standard

Randomness is a feature, but it must be explainable. Never use pure coin-flip events without a causal note.

Use three layers:

1. Type-weighted randomness: MBTI affects event probability and stress responses, but never decides a single fixed script.
2. Context-weighted randomness: the same event has different impact in romance, friendship, work, or family contexts.
3. Black-swan randomness: rare external events or timing shocks can sharply move the relationship.

Every event should expose:

• Event title.
• Event type: attraction, rupture, repair, pressure, ambiguity, external shock, boundary, opportunity.
• Variable deltas.
• Explanation chain: personality tendency + relationship context + current state + history.
• Resulting state movement, if any.

Boundaries and Limitations

• MBTI is used only as an entertainment and self-reflection framework. The sandbox shows possible trajectories; it does not provide scientific diagnosis, deterministic prediction, or professional psychological advice.
• Do not use it for abuse, coercion, self-harm, severe psychological distress, or legal, medical, or major financial decisions. In those cases, stop the forecast and redirect to qualified professional help.
• People nodes are limited to 2-20. Fewer than 2 people cannot form a relationship network; more than 20 is rejected.
• The deliverable is one self-contained HTML file. The skill does not build native apps, run a backend, or transmit user data anywhere.
• Conclusions use probabilistic language ("may", "tends to", "one plausible branch is"); avoid absolute claims about any person or type.

Failure Handling and Troubleshooting

• Malformed or invalid --seed-json: the generator rejects non-object input, invalid people arrays, more than 20 people, non-string names, and MBTI values outside the 16 supported types; it exits 1 before writing a page.
• Unwritable --output-dir or a path that is an existing file: the generator fails with a non-zero exit code instead of overwriting existing content.
• Missing template (assets/sandbox.html): the generator reports the missing file and exits 1 rather than writing an empty HTML file.
• Incomplete user inputs: the skill creates sensible defaults (starting from exactly 2 people) and keeps every field editable in the UI instead of blocking generation.
• Default output location: when --output-dir is omitted, the HTML is written into the current working directory, never to a hard-coded external path.

Resources

• assets/sandbox.html: self-contained HTML/CSS/JS dynamic sandbox template.
• scripts/generate_sandbox.py: copies the template into an output directory and optionally injects starter inputs.
• references/model.md: relationship state model, MBTI tendency matrix, event weighting rules, and safety language.

Use the script rather than rewriting the HTML from scratch unless the user asks for a custom implementation.

免费使用说明

当前版本按免费 Skill 发布，可直接使用。

直接按上方使用流程、脚本说明或输入输出格式执行。涉及公开网络、登录、验证码、权限、第三方页面或数据源限制时，仍按前文边界处理；不绕过平台规则，不用演示数据冒充真实结果。