Ruosong320
- 24 skills
- 0 followers
- 8 hours ago last updated
- ▌ Evidence Based Review · ruosong320 bundle依据真实受众或使用者反馈评估产品、内容或方案时使用；从外部反馈提炼指标，再用对象证据评分。普通代码审查或现成评分表不触发。
- ▌ Prompt Contract Review · ruosong320 bundle优化复杂提示词、多阶段 agent 输出或语义漂移时使用；核对指令、示例、schema、校验器与下游消费者是否表达同一契约。
- ▌ Experiment To Production · ruosong320 bundle将已验证实验、原型或新流水线接入正式流程时使用；核对输入、提示词、配置、状态和所有实际消费者，验证行为一致性。
- ▌
- ▌ Build Topic Learning Markmap · ruosong320 bundle将 input_datas/{topic} 中的本地资料和 CSV 链接整合为中文纯节点 markmap；适用于要求完整阅读来源并构建主题知识树的任务。
- ▌
- ▌
- ▌ Claude History Sync · ruosong320 bundle将用户指定的 Claude Code 或 Codex 历史导出、清洗并增量备份到已配置的 rclone remote；仅本地查看会话不触发上传。
- ▌
- ▌ Run Real Experiments · ruosong320 bundle运行方案试验、真实模型评测或多样例对比时使用；固定实验条件，调用真实入口，保留逐轮证据并检查跨样例改进。普通单元测试不触发。
- ▌ Initer · ruosong320 bundleClarify requirements before starting from-zero project builds or new feature work. Use when asked to create a complete new project, scaffold an app, design a system, add a new feature, introduce a new workflow, or make any change whose product, technical, UX, data, integration, security, deployment, or acceptance boundaries are unclear; generate a checklist from the user request, identify unclear boundaries, ask targeted multi-round questions, and begin implementation only after the checklist is sufficiently resolved.
- ▌ Ship It For Real · ruosong320Use only when the user explicitly invokes /ship-it-for-real or asks Claude to use ship-it-for-real for an authorized execution task that should be carried through implementation, recovery, and evidence-based completion. Use when there is NO pre-confirmed written plan — this skill supplies risk tiering, a retry budget, bounded monitoring, and terminal-state judgement. If a confirmed plan from real-solution-plan already exists and the job is to follow it faithfully, use ship-solution instead. Do not use for analysis-only requests, open-ended monitoring, or as permission for external or destructive actions.
- ▌ Understand First · ruosong320当用户刚描述完一套流程、规则、约束、判定标准、数据结构或工作方式，而 agent 即将据此实现、改造或批量执行时触发。触发词包括「我的思路是、流程是、规则如下、要求是、本质是、你理解一下、重新理解、证明你理解了、给几个例子看看、产出案例」。强制先证明理解：压缩复述规则 → 给出至少 3 个与用户举例不同领域的具体样例（含正例与反例）→ 指出规则内部的张力与边界疑点 → 给出实施骨架，等用户确认后才动手。禁止用用户自己举过的例子来证明理解。不接管需求澄清（归 initer）、方案设计（归 real-solution-plan）、只读分析（归 analyze-only）。
- ▌ Real Solution Plan · ruosong320当用户使用「给出方案、实现方案、解决方案、怎么实现、如何做、方案设计、技术方案、plan、solution plan」等词触发。产出详细的真实实现方案。要求基于真实项目环境和先验分析结论，结合网络检索的真实方法论和工程实践产出可执行方案。无先验分析时必须检索真实案例参考。不接受臆测、独立视角分析或无证据方案。
- ▌ Scoped Code Change · ruosong320Use when an agent needs to modify a local, specific, or partial area of a program, script, app, config, test, or code-adjacent text. Trigger for fixing one feature, changing a function/module/component, adjusting bounded behavior, or making a minimal patch. Derive the boundary from the request and repository, ask only about unresolved consequential decisions, then locate the smallest change point, match local style, patch, and smoke test immediately.
- ▌ Abstract First Coder · ruosong320Write Python code in the abstraction-first style used by knowledge_graph_xm: define domain data models, abstract interfaces, strategy classes, default implementations, pipeline orchestration, dependency injection, async contracts, and verification hooks before adding concrete behavior. Use when creating or modifying project code that should be extensible, replaceable, and readable by future agents.
- ▌ Recoder · ruosong320 bundleMaintain a durable project record for substantial work that spans sessions, changes architecture/interfaces/data, contains consequential decisions or failed approaches, or is explicitly requested for later handoff or methodology distillation. Capture both project state and the work path. Do not trigger for ordinary short or local changes, and do not require a record before such changes.
- ▌ Xm Chat · ruosong320 bundleUse for any Matrix chat work via the `xm chat` CLI: messaging (text/markdown/emote/notice/reply/edit), files/stickers/locations, history, real-time listening, daemon sessions, ACP-based AI virtual personas (`xm chat persona`), E2E encryption keys, appservice virtual identities, profiles/presence, room members, and event-triggered commands with context injection, batching, and agent session continuity. Covers all `xm chat` subcommands, event filter config (prefix/domain/regex matching, room/sender exclusion, @mention detection), and bot development. Trigger on: chat, messaging, Matrix, IM, bot, daemon session, chat trigger, virtual user, virtual persona, AI persona, chatbot, DM, room members, trigger pipeline, agent mode, XM_AGENT_SESSION, Monitor mode, XM_CHAT_USERNAME, XM_CHAT_PASSWORD, XM_CHAT_TOKEN, multi-device login, file send, sticker, location, emote, edit message, event filter, prefix filter, domain filter, persona debugging, router check, event tracing, or any room-based communication task.
- ▌ Analyze Only · ruosong320 bundle当用户使用「分析、仅分析、只分析、帮我分析、分析一下、评估、看一下、analyze、analysis、analyze only」等词触发。Perform objective, evidence-led analysis without changing any file, code, configuration, data, or persistent state. Use when the user asks to inspect, audit, review, diagnose, compare, validate, or assess a requirement, workflow, design, code path, document, dataset, result, decision, incident, or risk while requesting analysis only, no edits, no implementation, or a report before changes. Reconstruct the main workflow, locate the exact analysis point, examine upstream and downstream context, use a task-specific checklist and fit-for-purpose analytical methods, distinguish facts from inference and uncertainty, test alternative explanations, calibrate confidence, and report findings with traceable evidence. 强化要求：1）分析必须全面，输出前自检是否遗漏关键内容；2）无法得到准确结论、缺乏自信、无证据的分析不被接受，必须网络检索真实方法论和工程案例，结合多源参考给出分析，禁止臆测、偷懒、独立视角分析。
- ▌ Prompt Forge · ruosong320 bundle手动调用 /prompt-forge <被测物> <量化验收线>。把一个 skill / prompt / 流程迭代打磨到明确的量化验收标准（如「85 分，可用于真实场景拆解」）。循环：clean-room-eval 评测 → 定位共性问题根因 → 改写被测物 → 换全新案例集复测，直到达标或迭代预算耗尽。强制红线：禁止对着评测案例硬编码或加诱导性限定、禁止下调验收线、被证伪的体系整体弃用不打补丁、每轮必须换新案例集、连续两轮无改善即停止并如实上报。产出偏好选择题与完形填空式模板而非主观题。不自动触发（内含多轮子 agent 评测，开销大）。
- ▌ Ship Solution · ruosong320当用户使用「实现、执行方案、按方案做、开始实现、ship、implement solution」等词触发，或在 real-solution-plan 产出方案后用户确认时触发。严格按照 real-solution-plan 产出的方案执行实现，用真实案例验证，不接受补丁式修改和硬编码。小问题允许一次性修复，未达指标的问题需重新规划。用户确认方案后可选自动验证。前置硬要求：必须已有一份确认过的书面方案；如果没有方案、只是要把一件已授权的事推到有证据的终态（需要风险分级、重试预算、有界监控、终态判定），改用 ship-it-for-real。
- ▌ Clean Room Eval · ruosong320手动调用 /clean-room-eval <被测对象> [案例数]。用无污染子 agent 批量跑真实案例来评测一个 skill / prompt / 流程的真实能力，并归纳共性问题而非单点 bug。自己生成跨领域案例（不复用被测方给的案例）→ 开 N 个无对话历史、无其他 skill、无提示的干净子 agent 各跑一例 → 主 agent 并行联网检索人类基准做对照 → 只上报在多个案例中重复出现的共性问题，附命中率与逐例证据。不自动触发（会开大量子 agent，开销大，必须显式调用）。不做修复（修复归 prompt-forge），不做只读分析（归 analyze-only）。
- ▌ Mission Spliter · ruosong320 bundleSplit clarified requirements into phased missions AND accept each phase as it completes. Plans stage goals, deliverables, dependencies, task order, and a per-phase acceptance gate; then, after each phase is implemented, runs that gate against real evidence (Pass/Partial/Fail) before the next phase may start, and finally rolls phase verdicts up into one delivery acceptance. Use after requirements are clarified (by the user or Initer), before and during implementation of a project, major feature, workflow, integration, refactor, or multi-step task. If requirements are still unclear, hand back to clarification before splitting.
- ▌ Process Thinker · ruosong320流程思考者。手动调用 /process-thinker <目标>。三阶段：多轮弹窗摸底（身份/能力边界/环境条件/工具/成功指标）→ 倒推流程并逐节点联网检索真实工程案例佐证、预判卡点与规避方案 → 按固定模版输出 BLUF 报告并落盘。不自动触发。