s87343472/selfctx

# SelfCtx 切换模型不再从零开始。把你对 AI 说过的一切——工作方式、项目背景、行为规则——存成一组本地 Markdown 文件，粘贴进任何模型，立刻生效。 **[English ↓](#english)** ## 它解决什么问题 | 现状 | SelfCtx | |------|---------| | 切换模型要重新「调教」，花 30–50 分钟 | 粘贴一个文件，2 分钟恢复上下文 | | 上下文散落在各平台，找不齐 | 10 个维度，一个地方存所有 | | 平台 Memory 不能跨模型迁移 | 纯 Markdown，模型无关，本地掌控 | | 不知道哪些信息能给模型、哪些不能 | 4 个披露级别，按需选择 | ## 和同类方案的区别 - **vs 平台 Memory**：平台 Memory 是增量细节，SelfCtx 是结构化基线。两者互补，不是替代。 - **vs 手写 System Prompt**：System Prompt 只有偏好设置，SelfCtx 覆盖完整的人的上下文：叙事身份、能力专长、资产图谱、关系网络、项目状态…… - **vs Second Brain（Notion/Obsidian）**：Second Brain 存的是知识，SelfCtx 存的是「我是谁」和「AI 应该怎么对待我」。 ## 10 个维度 | # | 维度 | 更新频率 | |---|------|---------| | D1 | 叙事身份 — 我是谁 | 年级 | | D2 | **认知操作系统 — 对 AI 的强制要求** | 季度 | | D3 | 具身状态 — 健康与能量 | 月级 | | D4 | 情绪与心理（可选） | 月级 | | D5 | 能力与专长 | 季度 | | D6 | 资产图谱 — 产品/设备/代码库 | 季度 | | D7 | 关系网络 | 季度 | | D8 | 项目快照（3–12 个月窗口） | 月级 | | D9 | 兴趣与审美 | 年级 | | D10 | 世界观与志向 | 年级 | **D2 是核心。** 它存的不是偏好，是规则——粘贴进任何模型立刻强制执行： ai_hard_rules: forbidden: - "不说套话：禁止'很有意思''可能值得考虑'" - "不主动生成 docx/pptx，默认 markdown" required: - "必须表态：这能不能成立，依据是什么" - "方案给出 2-3 个路径并明确推荐" ## 4 个披露级别 | 级别 | Token | 包含 | 用途 | |------|-------|------|------| | L0 公开版 | < 500 | 身份标签 + 行为规则 | 随意粘贴，零风险 | | L1 工作版 | < 2,000 | + 项目 / 能力 / 工作关系 | 日常 AI 协作（推荐起点） | | L2 信任版 | < 5,000 | 全部维度，敏感信息脱敏 | 长期主力模型 | | L3 全量版 | 不限 | 原始数据 | 仅本地，永不外发 | ## 快速开始 1. 复制 prompts/P1_export.md 里的 prompt 2. 粘贴进已有上下文的模型 → 获得草稿 3. 检查 ⚠️ 标记，删除不想保留的内容 4. 填入 templates/pca_template.md 5. 生成 exports/L1_work.md 6. 粘贴进新模型 → 完成迁移 首次建立约 15–30 分钟。之后每次迁移约 2 分钟。 ## 文件结构 selfctx/ ├── schema.md # 10 个维度的完整字段定义 ├── templates/ │ └── pca_template.md # 空白模板，从这里开始 ├── prompts/ │ ├── P1_export.md # 从当前模型导出草稿 │ ├── P2_import.md # 导入新模型 │ ├── P3_import_quick.md # 快速导入（只加载行为规则） │ ├── P4_update.md # 更新单个维度 │ ├── P5_review.md # 季度检视 │ └── P6_level_select.md # 帮你选择披露级别 ├── exports/ # 你的导出文件（L0–L3） ├── my_pca/ # 你的原始 PCA 文件（按维度） └── skill/ # Claude 专属 Skill（可选） ## 模型兼容性（2026 年 6 月） | 模型 | L0 | L1 | L2 | |------|:--:|:--:|:--:| | Claude Opus 4.8 / Sonnet 4.6 | ✓ | ✓ | ✓ | | GPT-5.5 / GPT-5.4 | ✓ | ✓ | ✓ | | Gemini 3.1 Pro | ✓ | ✓ | ✓ | | Gemini 3 Flash | ✓ | ✓ | △ | | Grok 4 / 4.3 | ✓ | ✓ | △ | | DeepSeek V3.2 | ✓ | ✓ | △ | | Llama 4 Scout | ✓ | △ | △ | | Qwen 3.5 | ✓ | ✓ | △ | | Mistral Large | ✓ | ✓ | △ | ✓ 完全支持 · △ 基本可用，部分字段识别不够精准 ## Claude 用户 安装 `selfctx.skill` 获得增强：自动触发历史对话搜索、结构化导出、维度内联更新。 下载 `selfctx.skill` → Claude 设置中导入。需要 Claude 3.5+。 ## 协议 MIT · [Sagasu](https://sagasu.art) · [博客文章](https://sagasu.art) # SelfCtx Stop re-introducing yourself every time you switch models. Store everything you've ever told an AI — your work style, projects, behavior rules — as a set of local Markdown files. Paste into any model. Works instantly. ## What problem it solves | Before | With SelfCtx | |--------|-------------| | 30–50 min to re-onboard a new model | Paste one file, context restored in 2 min | | Context scattered across platforms | 10 dimensions, one place for everything | | Platform memory can't cross models | Plain Markdown, model-agnostic, locally owned | | Unsure what's safe to share | 4 disclosure levels, you decide what goes where | ## How it compares - **vs Platform Memory**: Memory handles recent incremental details. SelfCtx handles the stable structural baseline. Complementary, not competing. - **vs System Prompts**: System prompts carry preferences. SelfCtx carries the full picture: narrative identity, expertise, asset map, relationships, active projects… - **vs Second Brain (Notion/Obsidian)**: Second Brain stores knowledge. SelfCtx stores *who you are* and *how AI should treat you*. ## 10 Dimensions | # | Dimension | Update Freq | |---|-----------|-------------| | D1 | Narrative Identity — who I am | Yearly | | D2 | **Cognitive OS — hard rules for AI** | Quarterly | | D3 | Embodied State — health & energy | Monthly | | D4 | Emotional & Psychological (optional) | Monthly | | D5 | Capabilities & Expertise | Quarterly | | D6 | Asset Map — products / devices / repos | Quarterly | | D7 | Relationship Network | Quarterly | | D8 | Project Snapshot (3–12 month window) | Monthly | | D9 | Interests & Aesthetics | Yearly | | D10 | Worldview & Aspirations | Yearly | **D2 is the core.** It stores not preferences but enforced rules — active the moment you paste them into any model: ai_hard_rules: forbidden: - "No filler: never say 'that's interesting' or 'you might want to consider'" - "Never generate docx/pptx unprompted — default to markdown" required: - "Take a position: state whether something will work and why" - "Always offer 2–3 distinct paths with a clear recommendation" ## 4 Disclosure Levels | Level | Tokens | Contents | Use Case | |-------|--------|----------|---------| | L0 Public | < 500 | Identity tags + behavior rules | Share freely, zero risk | | L1 Work | < 2,000 | + Projects / expertise / work relationships | Daily AI collaboration (recommended start) | | L2 Trusted | < 5,000 | All dimensions, sensitive fields redacted | Long-term primary model | | L3 Full | Unlimited | Raw data | Local only, never share | ## Quick Start 1. Copy the prompt from prompts/P1_export.md 2. Paste into a model that already knows you → get a draft 3. Review ⚠️ markers, remove anything you don't want 4. Fill in templates/pca_template.md 5. Generate exports/L1_work.md 6. Paste into a new model → done First setup: ~15–30 min. Every migration after: ~2 min. ## File Structure selfctx/ ├── schema.md # Full field definitions for all 10 dimensions ├── templates/ │ └── pca_template.md # Blank template — start here ├── prompts/ │ ├── P1_export.md # Export draft from current model │ ├── P2_import.md # Import into new model │ ├── P3_import_quick.md # Quick import (behavior rules only) │ ├── P4_update.md # Update a single dimension │ ├── P5_review.md # Quarterly review │ └── P6_level_select.md # Help choosing the right level ├── exports/ # Your export files (L0–L3) ├── my_pca/ # Your raw PCA files (one per dimension) └── skill/ # Claude-specific Skill (optional) ## Model Compatibility (June 2026) | Model | L0 | L1 | L2 | |-------|:--:|:--:|:--:| | Claude Opus 4.8 / Sonnet 4.6 | ✓ | ✓ | ✓ | | GPT-5.5 / GPT-5.4 | ✓ | ✓ | ✓ | | Gemini 3.1 Pro | ✓ | ✓ | ✓ | | Gemini 3 Flash | ✓ | ✓ | △ | | Grok 4 / 4.3 | ✓ | ✓ | △ | | DeepSeek V3.2 | ✓ | ✓ | △ | | Llama 4 Scout | ✓ | △ | △ | | Qwen 3.5 | ✓ | ✓ | △ | | Mistral Large | ✓ | ✓ | △ | ✓ Full support · △ Works, some fields less accurately parsed ## Claude Users Install `selfctx.skill` for enhanced features: automatic conversation history search, structured export, inline dimension updates. Download `selfctx.skill` → import in Claude settings. Requires Claude 3.5+. ## License MIT · [Sagasu](https://sagasu.art) · [Blog post](https://sagasu.art)

标签：Markdown, Ruby, 上下文管理, 人工智能, 提示词工程, 效率工具, 用户模式Hook绕过, 知识库, 策略决策点, 防御加固