hermes-config-kit

Set up Alembic database migrations for FastAPI projects using async SQLAlchemy (asyncpg). Covers env.py configuration with async engine, autogenerate handling, greenfield vs brownfield migration strategies, and env var integration.

使用 CloakBrowser 绕过抖音反爬，获取真实视频 MP4 下载链接

Wan2.2 FLF2V (First/Last Frame to Video) 管线 — 生成首尾帧之间的插值视频，支持电影级镜头过渡

Use CloakBrowser (anti-detection Chromium) to access protected Chinese web platforms (Douyin/抖音, etc.) and extract structured data — intercepting API responses, extracting video/media URLs, and downloading content from sites that block standard Playwright/browser automation.

Install and diagnose Playwright-based packages (Playwright, CloakBrowser, etc.) on Linux/WSL — handles externally-managed Python, system deps, proxy-based binary downloads, and runtime troubleshooting

Exploratory QA of web apps: find bugs, evidence, reports.

llama.cpp local GGUF inference + HF Hub model discovery.

Build an end-to-end ML emotion classification pipeline from raw tabular/biosignal data — heuristic label generation, feature engineering, RandomForest training, Flask API, and SQL schema.

将 bili-fe-workflow 方法论移植到 Hermes 的标准化开发工作流。用 AGENTS.md + Memory + Skills 三件套替代 .workflow 知识库，实现从需求到提测的 AI 全链路自动化。

Blueprint for building full-stack ML web app course projects (BERT + Flask + Vue) in WSL — handles /mnt/ mount restrictions, spec-driven development, model training, and course documentation.

Build a vertical scroll-snap feed UI (TikTok/Douyin-style) in Vue 3 — touch drag-to-snap, keyboard navigation, mixed content types, Canvas-based placeholder videos, and pixel-perfect matching from screenshot references.

Build an AstrBot plugin — covers fork-existing vs from-scratch decision, 4-quadrant research gate, doc reading order, decorator mechanics, schema field patterns, ad-hoc verification, CI preflight script. Use when user asks to "做一个 AstrBot 插件" or "集成 X 到 AstrBot" or describes a bot feature.

ComfyUI API 调用、工作流构建、模型管理和提示词编写 — 针对实验室 4×A10 服务器

在 WSL/Linux 上通过 uv 部署 AstrBot QQ 机器人，对接已有 NapCatQQ

Install pre-built external AI agent skills from GitHub (agentskills.io / Claude Code / OpenClaw ecosystem) directly into Hermes Agent's skills directory. Two methods: (1) `npx skills add` (agentskills.io CLI) — fast but may fail on format, (2) `git clone` fallback — always works, preserves subdirectory skills and assets. Use when the user says "装一下这个skill", "安装这个技能", "装个XX", "install this skill", or points to a GitHub repo that contains a SKILL.md.

Use ServerPackCreator (Griefed) CLI to convert a Minecraft client modpack (CurseForge .zip or Modrinth .mrpack) into a runnable server pack. Java 21, Arch/WSL friendly. Outputs zip + flat directory with start.sh/.bat/.ps1 + server.properties. Use when user has a client modpack and needs a server-side jar/scripts.

Diagnose Forge 1.18-1.20 modpack clients (PCL2/HMCL/MultiMC) that hang at startup — no crash, no main menu, splash freezes. Trigger on "卡住", "hang", "stuck loading", "won't enter main menu", "启动卡死", "整合包启动不动".

Convert a client Minecraft modpack (CurseForge / Modrinth / Modrinth .mrpack) into a server pack ready to host with Forge / NeoForge / Fabric / Quilt. Use when the user has a modpack zip or modpack instance and wants to run it as a multiplayer server — covers ServerPackCreator (SPC) CLI on WSL/Linux, the Modrinth-vs-CurseForge format difference, the three conf fields cgen leaves blank, the inclusions format for Modrinth overrides/, and the Windows PowerShell SMB path for moving the zip out of WSL.

Use when a user asks whether an AI project, agent, or workflow can autonomously make real money for them — including "AI 赚钱", "AI 被动收入", "AI 自动交易", "let AI make money for me", "AI side hustle", "passive income with AI", "AI 数字人赚钱", "AI 量化 自动下单". Evaluates claims through a 3-axis filter (real revenue path × regulatory friction × disclosure honesty) and produces a routed verdict (real-and-runnable / jurisdiction-blocked / research-grade / scam). Does NOT itself attempt to make money; identity-bound actions (registration, KYC, payments, real brokerage/exchange accounts) require explicit per-call user authorization.

Token-Oriented Object Notation — encode/decode JSON to compact TOON format (~40% token reduction for uniform arrays). TRIGGER when user: (a) pastes JSON inline (file or `cat X.json | ...`), (b) mentions "compress", "shrink", "encode JSON", "TOON", "tokens", "context window", "省 token", (c) asks to "analyze / summarize / inspect" a JSON dataset, (d) processes an API response with >2KB JSON output, (e) reads JSON files into context for processing. Always check: `cat X.json | ~/.local/bin/toon` is safe to run (non-JSON passes through). If output goes to file/DB/written verbatim, skip TOON. If nested non-uniform, JSON may be smaller — measure both.

WSL 开发环境配置 — NTFS venv 权限、Clash 代理访问、DeepSeek ReAct 模式、PaddlePaddle 安装、>60s 长任务的 Hermes sandbox 限制、GitHub release 大文件下载

AI漫剧/短剧管线选型评估 — 当用户想切换风格（写实→动漫/2D/3D）、评估生产路线、或不确定该走哪条管线时，通过系统性的资产盘点→选项调研→权衡对比→推荐，辅助决策。

Flux 2 Klein 原生参考图角色一致性生成 — 使用 ReferenceLatent 节点零样本保持角色身份，替代旧有的 prompt engineering 风格锚定方案。适用于 AI 短剧管线中的角色资产生成。

用 Hermes 的 file 工具读写文件时避免数据损坏——read_file 会注入行号，不要直接拿去 write_file

Package a finished full-stack ML course project (Flask + Vue + PyTorch model) into a zero-config, single-click-runnable deliverable for teacher grading. Covers: pre-building Vue frontend, modifying Flask to serve static frontend, generating frozen requirements.txt, creating platform launcher scripts, and zipping for delivery.

人民公仆（Publieople）的思维框架与表达方式。基于49个Notion页面的深度蒸馏， 提炼5个核心心智模型、10条决策启发式和完整的表达DNA。 用途：作为思维顾问，用人民公仆的视角分析技术问题、评估工具、优化工作流、组织知识、做效率决策。 配套文件：work.md（方法体系与工作流）、persona.md（思维框架与表达DNA）、meta.json（结构化元数据）。 当用户提到「用Publieople的视角」「切换回自己的视角」「切换到人民公仆」时使用。 进化模式：「追加」补充素材、「纠正」修正内容、「/update-skill publieople」全面更新。

人民公仆（Publieople）的思维框架与表达方式。基于49个Notion页面的深度蒸馏， 提炼5个核心心智模型、10条决策启发式和完整的表达DNA。 用途：作为思维顾问，用人民公仆的视角分析技术问题、评估工具、优化工作流、组织知识、做效率决策。 当用户提到「用Publieople的视角」「切换回自己的视角」「切换到人民公仆」时使用。

AstrBot 运营维护 — 进程管理、插件热重载、人格系统、知识库集成、常见故障排查

从 Wikimedia Commons 抓取自由许可的真实图片素材(产品/物品/场景)。用于 U 盘素材准备、文档配图等需要真实照片但无 API key 的场景。

将 AstrBot 连接到远程 ComfyUI 服务器 — SSH 隧道、systemd 自启、插件选型、工作流配置

Diagnose why a Hermes Agent setting/flag/env var is or isn't taking effect. Use when the user reports "I set X, reloaded, still Y happens" against ANY Hermes surface (CLI prompt_toolkit, TUI Ink overlay, gateway, Electron desktop, dashboard PTY, env vars from config.yaml). Triggers: "why does /new still prompt", "I set approvals.X = false but it still asks", "the config flag isn't honored", "env var not reaching the subprocess", "TUI vs CLI setting", "which config does this read". Hermes is a polyglot stack: a single user-facing behavior (e.g. "destructive command confirmation") is gated in MULTIPLE independent surfaces, each with its own config key, env var, or read path. The fix that "didn't work" is almost always the right toggle applied to the wrong surface.

Hermes ships FOUR user-facing surfaces (CLI prompt_toolkit, Ink TUI in Python, Node ui-tui bundle, Electron desktop, web dashboard, messaging gateway). Each has its OWN gate for destructive operations, env var inheritance, and config persistence. A `config.yaml` flip in one does NOT silence the modal in another. Use when the user says "I turned off X but it still pops up", "why doesn't this config take effect", "I set HERMES_*_NO_CONFIRM=1 but the dialog is back", or asks to disable/inspect any cross-cutting runtime behavior (confirmations, telemetry, key bindings, approval gates, resume semantics).

Reverse-engineer user-visible Hermes behavior by reading `~/.hermes/hermes-agent/` source — locate the source of a slash command, config key, confirmation modal, or UI element, then explain root cause and (when applicable) the bypass. Use when a user asks "why does X happen", "how do I turn off Y", "where is Z implemented", or "what's the config key for W" and the answer isn't in the docs site.

Auto-launch a Windows desktop script or program with admin privileges at user logon. Covers the three viable mechanisms (Task Scheduler RunLevel=Highest, startup-folder shortcut, AHK self-elevate), the .lnk RunAs-flag pitfall, and how 3rd-party launchers (Quicker, Power Automate, etc.) fit into the picture. Use when a user reports "works when I double-click it but doesn't start at boot", "startup shortcut silently fails", or wants to add a logon-time admin task.

Prepare for and drill Office practical exams (Word + Excel + Access) — BTEC / HND / ITAS2-style supervised open-book assessments with pre-exam materials gathering, themed tasks (typically 20-30 task booklet), file-format restrictions, and no internet. Use when user asks for exam practice, mentions ITAS2 / HP12 48 / BTEC IT / HND Computing practical, or hands over an Assessment booklet + Required Materials notice.

Bulk-generate Office deliverables (docx/xlsx/pdf) via python-docx + openpyxl + LibreOffice headless when user is out of time on an exam. Linux-friendly. Use when user asks to bulk-generate or has <30 min left.

AstrBot 插件开发 — 骨架、装饰器、配置 schema、LLM 工具、注册/发布、热重载验证。Fork 优先于从零写,文档先于代码。

ITAS2 (HP12 48) Re-Assessment 1 考试辅助 — Vibe Digital Experience Centre 主题，27 Task 操作手册 + 真实素材表结构 + 考场策略。考试当天 (2026-07-10) 调用。

大学生活质量指北 — 从 colleges.chat 查询全国高校的真实学生反馈（宿舍、空调、独立卫浴、断电断网、食堂等生活质量信息）

Maintain a GitHub fork of an active upstream open-source project — diagnose sync failures, recover from long-term divergence, preserve local personalisation (config files, themes, assets), and re-arm auto-sync. Covers NotionNext, Next.js starters, Astro themes, Vercel templates, plugin frameworks (AstrBot, etc.) and any fork that upstream ships updates you want while keeping your overlay. Also covers the **first-time fork + local personalise** flow when the user wants a personal fork they intend to push commits to, not just sync with upstream.