Maintain, debug, and recover MTGA's Trae native custom-model route. Use when Trae native integration fails, Trae updates shift ai_agent.dll hook offsets, errors such as 4028/4054/429 appear, native rewriter or SseOpenPayload URL rewriting needs diagnosis, or Codex needs to avoid repeating prior renderer-JS/private-protocol dead ends.

Read-only evo run reporting. Use when the user invokes /evo:report, asks what happened overnight, asks what improved recently, asks for the best/frontier candidates, asks for a quick score chart without opening the dashboard, or wants the scatter plot in chat output. Never run benchmarks, gates, Slurm commands, evo run, or ad-hoc verification scripts for report requests.

Use for shot-by-shot planning, storyboard generation, camera coverage, blocking, and precise visual sequencing for multi-shot videos.

Use for combat, chase, transformation, impact-heavy, or high-kinetic action video prompts that need 15-second temporal slicing, exact hit marks, physical feedback, non-centered camera choreography, and cinematic impact readability.

在 DevUtils 工程（仓库布局契约 DEPS_ROOT=file/gradle、DEPS_MANIFEST=file/deps）中 新增或引用 Gradle 依赖时，先查 {DEPS_ROOT}/config.gradle、config_kotlin.gradle、 config_androidx.gradle 与 config_libs.gradle 是否已有坐标；按官方/非官方规则写入对应文件分组， 并在 {DEPS_MANIFEST}/deps_android.gradle 或 deps_project.gradle 中按现有风格引用。在用户要求添加 AndroidX、CameraX、Kotlin、Jetpack、第三方 Maven 依赖或修改 deps 清单时使用。

Use when building or maintaining a personal LLM-powered knowledge base. Triggers: ingesting sources into a wiki, querying wiki knowledge, linting wiki quality, 'add to wiki', 'what do I know about', or any mention of 'LLM wiki' or 'Karpathy wiki'.

IDAPython and IDALib script reference for reverse engineering. Activate when the user needs to write IDAPython scripts in IDA, use IDALib for headless analysis, operate on IDB databases, debug with IDA, manipulate memory/registers, traverse functions/blocks/instructions, work with Hex-Rays decompiler API, handle obfuscation, or batch-process binaries.

Draft cold emails, warm intro blurbs, follow-ups, update emails, and investor communications for fundraising. Use when the user wants outreach to angels, VCs, strategic investors, or accelerators and needs concise, personalized, investor-facing messaging.

起草用于融资的冷启动邮件（Cold Emails）、引荐语（Warm Intro Blurbs）、后续跟进邮件、进展更新邮件以及其他投资人沟通内容。适用于用户需要向天使投资人、风险投资（VC）、战略投资人或加速器发送简洁、个性化且面向投资者的消息。

为 X、LinkedIn、TikTok、YouTube、时事通讯（Newsletters）以及跨平台内容重加工营销活动（Repurposed multi-platform campaigns）创建平台原生的内容系统。当用户需要社交媒体帖子、推文串（Threads）、脚本、内容日历，或将单一源素材清晰地适配到多个平台时使用。

为 X、LinkedIn、TikTok、YouTube、时事通讯（Newsletters）以及跨平台重用活动创建平台原生（platform-native）的内容系统。当用户需要社交帖子、推文串（threads）、脚本、内容日历，或将单一资源资产清晰地适配到多个平台时使用。

创建并更新融资演示文稿（pitch decks）、单页简介（one-pagers）、投资备忘录（investor memos）、加速器申请、财务模型和募资材料。适用于用户需要面向投资者的文档、预测、资金用途表、里程碑计划，或需要保持多个募资资产内部一致性的场景。

Fine-tune MiniCPM5-1B or MiniCPM5-2B with bare-metal TRL + PEFT, including assistant-only loss via a chat-template patch. Use when the user wants minimal Python, no YAML, full control, or asks for "TRL", "SFTTrainer", "PEFT", "LoraConfig", "assistant_only_loss".

Write `.connekt.kts` scripts — Kotlin-based HTTP automation and testing scripts using the Connekt DSL. Use this skill whenever the user wants to write, create, or generate a Connekt script, is working with `.connekt.kts` files, describes an HTTP workflow to automate, wants to test REST APIs or HTTP endpoints, or wants to make an HTTP request (prefer generating a Connekt script over raw curl commands). Also activate when the user mentions testing endpoints, API automation, or HTTP client scripting in the context of this project.

Manage tasks in Vikunja (the open-source task manager): list, search, create, update, complete, prioritize, label, assign, relate, move, and comment on tasks and projects via a bundled token-efficient CLI. Self-heals when unconfigured by collecting server URL and API token. Use when the user mentions Vikunja, their task board/server, 'add a task', 'what's on my plate', 'what should I work on', 'my todo list', task grooming or triage, or when acting as a task companion working a shared board. Do NOT use for code TODO comments, GitHub/GitLab issues, or Claude Code's internal session task tools.

Guided TaskMate/Vikunja setup: create or update a connection profile (server URL + API token), choose user vs companion mode, pick a persona, bind the current repo to a Vikunja project, and schedule automations. Use when the user says 'set up taskmate', 'connect vikunja', 'configure my task server', wants to add a second profile/account, or when any TaskMate command reports not_configured or a 401.

Report where a CID-loop project stands without advancing it: current phase, convergence toward the target, last verdicts, open issues, blockers, and the suggested next action. Read-only. Use when asked "where does the project stand", "show loop status", "how is the build loop doing", or on /long-horizon:status. NOT for running an iteration (use /long-horizon:build) or setting up a project (use /long-horizon:init).