AI Agent & Automation Workflows

Unifying control-systems metalayer for LLM-as-controller agent development. Bootstrap any repository with typed plant/action/trace schemas, safety shield conventions, multi-rate loop hierarchy, EGRI-compatible evaluator interfaces, and the full consciousness stack (governance + knowledge graph + episodic memory). Use when: (1) setting up agentic control primitives in a new or existing project, (2) designing LLM-as-controller architectures with safety shields and typed directives, (3) wiring EGRI/autoany loops for controller or artifact improvement, (4) bootstrapping the consciousness stack (control-metalayer + knowledge-graph + conversation bridge), (5) integrating symphony-style orchestration patterns, (6) defining plant interfaces, state estimators, or world models for agent-controlled systems, (7) user says "control kernel", "agentic control", "safety shield", "plant interface", "control metalayer", "agent controller", "multi-rate loop", "LLM control law".

Provision and harden a fresh Linux VPS into an autonomous-agent dev host using the capability-preserving model — the box IS the sandbox: full agent autonomy inside it (non-root sudo user, open egress, YOLO mode, any toolchain), contained by ISOLATION (non-root + VPN-only access + snapshot + no-long-lived-secrets), not by restricting the agent. Composes an idempotent provisioning script, a staged firewall/VPN lockout-safe sequence, and a deterministic verification gate. USE WHEN setting up a new VPS/cloud box for Claude Code / agentic development, hardening a machine to run coding agents unattended, "set up another machine like srv1692698", provisioning an agent dev host, or auditing whether a box meets the agentic-host security invariants. NOT FOR generic server hardening with no agent on it (that's standard CIS/Lynis), desktop/laptop setup, or Kubernetes/multi-tenant clusters. Triggers on "set up a vps for agents", "agent dev host", "harden a box for autonomous agents", "new agent machine", "provision agentic vps", "agentic-vps".

Estimate optimal waiting time for Alkosto's "every 25/50 customers" promotion using either checkout-flow observations or winner announcement timestamps.

BroomVA trademark web styling system — Arcan Glass design language for Next.js + Tailwind v4 + shadcn/ui projects. Use when creating or styling BroomVA web interfaces, implementing glass/frosted UI effects, setting up brand-consistent dark-first themes, or integrating ArcanBG brand tokens (AI Blue 0066FF, Web3 Green 00CC66) into web applications. Provides a drop-in globals.css, glass component patterns, OKLCh color system, and shadcn/ui compatibility.

Distilled architecture & design principles for building self-service developer platforms, control-plane / data-plane separation, and edge-centralized cross-cutting concerns — drawn from a senior platform engineer's 8-year retrospective on building Atlassian's Envoy-based load-balancing platform (Open Service Broker + Sovereign xDS control plane + AWS infra + sidecar model). Covers the three IaC layers (image / infra / runtime), the template+context pattern for dynamic proxy configuration, forced-migration as an adoption tool, long-term maintenance discipline (churn as a smell, build vs maintain as distinct skills), and the non-technical engineering skills (diplomacy, mentoring vs teaching). Use when: (1) designing a self-service internal developer platform, (2) deciding between native proxy features vs sidecars, (3) introducing or migrating to Envoy / xDS / service mesh, (4) implementing the Open Service Broker spec or any resource-provisioning API, (5) abstracting cloud complexity (AWS CloudFormation, Packer, SaltStack/Ansible/Puppet) behind a developer-friendly API, (6) centralizing cross-cutting concerns (auth, authz, rate limiting, DDoS, access logs) at the edge, (7) planning a fleet-wide migration to new infrastructure, (8) auditing a long-lived codebase for churn / coupling / maintenance burden, (9) onboarding engineers to an existing platform, (10) deciding when to mentor vs train, (11) user says "platform engineering", "control plane", "data plane", "xDS", "Envoy", "service broker", "edge compute", "sidecar", "cross-cutting concerns", "self-service", "internal developer platform", "IDP", "platform migration", "code churn", "long-term maintenance", "mentoring vs teaching".

Microsoft BitNet — 1-bit LLM setup, inference, and benchmarking on CPU. Automates the full workflow: clone bitnet.cpp, create conda env, download GGUF models from HuggingFace, build optimized ternary kernels, and run inference. Supports official Microsoft models (2B) and community models (0.7B-10B). Use when: (1) setting up BitNet/bitnet.cpp for local CPU inference, (2) downloading and running 1-bit/ternary LLMs, (3) benchmarking BitNet vs full-precision models, (4) building edge/agentic inference pipelines without GPU, (5) converting HuggingFace models to GGUF for bitnet.cpp. Triggers on: 'bitnet', '1-bit llm', '1.58-bit', 'ternary model', 'ternary weights', 'edge inference', 'cpu inference', 'bitnet.cpp', 'bitlinear', 'no gpu inference'.

Quick capture of raw thoughts with intelligent domain classification and competitive intelligence…

Produce high-retention, dopamine-aware video content using brainrot editing techniques — fast cuts, word-by-word captions, sound design, visual velocity, pattern interrupts — all in service of genuinely valuable content. The Fireship model: brainrot pacing meets real substance. Built on Remotion with Imagen/Veo for assets. Use when: (1) creating edutainment content that competes with brainrot for attention, (2) making technical content accessible to younger audiences, (3) producing TikTok/Reels/Shorts that teach while entertaining, (4) applying engagement psychology ethically, (5) adding brainrot energy to a launch video or product demo. Triggers on: 'brainrot', 'edutainment', 'fast-paced video', 'tiktok style', 'shorts style', 'dopamine video', 'high retention', 'fireship style', 'brainrot for good'.

Brand icon and visual identity management for BroomVA projects. Generates favicons, app icons, Open Graph images, and social media avatars from brand assets. Maintains consistent visual identity across web, mobile, and social platforms. Use when: (1) generating favicon sets (ICO, PNG, SVG) for a web project, (2) creating app icons for iOS/Android, (3) generating Open Graph and Twitter Card images, (4) producing social media profile/banner images, (5) ensuring brand consistency across platforms.

CLI for broomva.tech — manage prompts, skills, and context from the terminal. Pull/push prompts, browse the skills roster, and query project conventions.

bstack primer — the Broomva Stack: twenty irreducible automation primitives (P1–P20) that turn an agent-driven workspace into a self-operating system. This is the AGENT-READABLE PRIMER (primitive contract + short-name index + when each fires), NOT the CLI. The bstack CLI + governance substrate (bin/, doctor, templates, hooks) is a separate product installed by clone + bootstrap (git clone https://github.com/broomva/bstack && ./bin/bstack bootstrap) — do NOT npx skills add broomva/bstack (it drops bin/scripts, the repo-root SKILL.md case of vercel-labs/skills1523). Use this primer to reason with the primitives: Bridge (P1), Gate (P2), Tickets (P3), Pipeline (P4), Fanout (P5), Bookkeeping (P6), Freshness (P7), Janitor (P8), Wait (P9), Hygiene (P10), Empirical (P11), Persist (P12), Dream (P13), Dep-Chain (P14), Snapshot (P15), Crystallize (P16), Lens (P17), Audience (P18), Orchestrate (P19), Cross-Review (P20). Triggers on 'bstack', 'Broomva Stack', 'the primitives', 'P1'..'P20', 'primitive contract', 'bstack bootstrap', 'governance metalayer', 'self-operating workspace', 'how do I install bstack', 'what does Pn mean'.

Agentic robotics with CaP-X — LLM-driven robot manipulation via code generation. Use when: (1) Setting up CaP-X / CaP-Gym environments for robot manipulation benchmarks, (2) Running CaP-Bench evaluations across LLMs/VLMs on robotic tasks, (3) Building or extending CaP-Agent0 agentic harnesses (skill libraries, visual differencing, parallel reasoning), (4) Training robot coding agents with CaP-RL (GRPO on code generation), (5) Developing perception APIs (SAM3, Molmo, depth, point clouds) or control APIs (IK solvers, grasp planners), (6) Sim-to-real transfer for Franka Panda, R1Pro humanoid, or other robot platforms, (7) Designing auto-synthesized skill libraries for physical manipulation (Voyager-style), (8) Integrating agentic robotics with Life Agent OS (Arcan orchestration, Spaces agent networking, Lago persistence), (9) Any task involving LLM-based robot control, manipulation benchmarks, robotic code synthesis, or embodied AI agents.

Ingest-and-integrate an artifact someone points at with a terse, deliberately under-articulated directive — "check this out", "lets research this", "look into this", "wdyt" — followed by a URL, repo, paper, file, image, or pasted document. Infers the actual request from a fully-contextualized frame (who is asking + what they're working on + what's already known) WITHOUT bouncing a clarifying question back, then runs the pipeline: contextualize → deep research (verify every source) → analyze → connect → document the finding → suggest ranked next steps. The artifact ends up metabolized into knowledge and action, not left as a one-paragraph summary. Composes existing research/search/memory tools; it does not reimplement them. USE WHEN: check this out, checkit, /checkit, lets research this, research this, look into this, dig into this, wdyt, what do you think of this, found this, take a look + an artifact; or whenever an artifact is shared with an underspecified ask. NOT FOR: retrospective "what have I been doing repeatedly" (a discovery/look-back task); fully-specified artifact asks ("summarize this PDF in 3 bullets", "fix the bug in this file" — answer directly). Triggers on those phrases + a shared artifact.

Set up Claude Code messaging channels for Telegram and Discord — bot creation, plugin installation, token configuration, access control (pairing, allowlists, guild channels), and troubleshooting. Use when the user asks to connect Claude Code to Telegram or Discord, configure a messaging bot, set up channel access, enable guild channels or threads, troubleshoot MCP server failures, or manage channel permissions. Triggers on "telegram channel", "discord channel", "claude code channels", "messaging bot", "set up telegram", "set up discord", "pairing code", "discord bot", "telegram bot", "channel access", "--channels".

Per-channel remote sessions for Claude Code via Discord and Telegram — each channel, thread, or chat gets its own isolated Claude Code session via tmux, with per-channel access control, project-specific workdirs, and automatic CLAUDE.md chain loading. Includes session managers, watchdog daemons (auto-respawn, auto-discover, stale cleanup), thread context injection, session resume, workdir mapping, and launchd boot persistence. Use when: (1) setting up per-channel Discord/Telegram sessions for Claude Code, (2) managing multiple sessions across messaging channels, (3) auto-discovering new channels or threads, (4) spawning thread sessions with parent conversation context, (5) keeping remote agent sessions alive, (6) cleaning up stale sessions, (7) setting up Telegram per-chat sessions. Triggers on 'remote sessions', 'discord sessions', 'telegram sessions', 'per-channel discord', 'discord watchdog', 'telegram watchdog', 'spawn discord session', 'claude remote', 'channel session', 'thread session'.

Orchestrate Google Colab Pro/Pro+ GPU instances as remote training backends via SSH. Compounds /agent-browser (to launch Colab sessions and install colab-ssh) with SSH (to operate the runtime remotely). Use when: (1) launching a Colab notebook for GPU training, (2) running training jobs on Colab from the local terminal, (3) transferring datasets/checkpoints to/from Colab, (4) monitoring GPU utilization on a Colab instance, (5) integrating Colab GPU compute with /autoany EGRI optimization loops, (6) reconnecting to a Colab session after timeout. Triggers on: "colab", "colab-remote", "colab ssh", "colab training", "remote GPU", "colab pro", "train on colab", "google colab".

Knowledge engine over the Colombian Truth Commission (CEV) final report "Hay Futuro Si Hay Verdad" (2022) — the findings, statistics, armed-actor responsibilities, differential harms, coined lexicon, and the 67-recommendation roadmap for non-repetition (no repetición). Uses a kg/LLM-wiki retrieval architecture (substrate-canonical knowledge pages + a dense catalog projection for tier-1 routing + body-grep tier-2 + the agent as query engine) plus data-backed tools, including an align check that scores a proposed policy/action against the report's recommendation roadmap. USE WHEN reasoning about the Colombian armed conflict, peace process, transitional justice, victims, reconciliation, paramilitarism, the FARC-EP/ELN, narcotrafficking and drug policy, land/agrarian questions, the enemigo interno, falsos positivos, the Unión Patriótica genocide, social-leader killings, differential harms (ethnic peoples, women and LGBTIQ+, children, exiles, campesinado), territorial/human security, memory, or non-repetition; when designing, evaluating or critiquing a peacebuilding/policy intervention against the CEV evidence; when asked what the report found or recommends on a topic; or when you need a sourced conflict statistic, responsibility share, or coined concept. NOT FOR downloading/ingesting the report (that was a one-off pipeline); NOT FOR live news on the current Colombian government; NOT FOR individual criminal responsibility (that is the JEP's role — this skill carries the CEV's COLLECTIVE ethical-political responsibility). Triggers on "comisión de la verdad", "truth commission", "hay futuro si hay verdad", "Colombian conflict", "non-repetition", "no repetición", "paz grande", "transitional justice Colombia", "falsos positivos", "paramilitarismo", "JEP/CEV".

Agent→human teach-to-mastery loop. Turns the agent into a wise, effective teacher whose single objective is that YOU deeply understand a body of work — the problem and why it existed, the solution and why it was resolved that way (design decisions, edge cases), and the broader context (why it matters, what it impacts). Stage-gated (confirm mastery before advancing), active-recall driven (you restate first; the agent fills gaps; quizzes via AskUserQuestion), and goal-bounded (the session does not end until your understanding is verified). Default subject = the current session/diff; pass a PR number, file, or subsystem to teach that instead. The pedagogical INVERSE of grill-me: grill-me stress-tests YOUR forward plan; comprehend transfers mastery of EXISTING work to you. Distinct from handoff (agent→agent narrative) and Bridge/Bookkeeping (agent→KG). Use when: (1) you want to deeply understand what was just built, (2) onboarding yourself or a teammate onto a PR / file / subsystem, (3) post-mortem learning after a complex session, (4) you say "teach me this", "help me understand", "walk me through", "make sure I get this", "I want to actually understand X". Triggers on "comprehend", "/comprehend", "teach me", "help me understand", "walk me through", "make me understand", "explain this so I get it", "make sure I understand", "onboard me", "I want to learn how this works".

Browser Orchestration — Playwright-driven generation at scale. Claude Code runs your entire content pipeline in the browser. Saves authentication sessions per tool, executes batch generation using compiled identity, and organizes output automatically. Supports Higgsfield, Weavy, Artlist, and extensible to any web-based generation tool. Triggers on: 'autopilot', 'browser automation', 'batch generate', 'setup tool', 'automated generation'.

Cinematic Generation Layer — the taste and technique library for premium AI content. Codifies start-frame doctrine, camera style vocabulary (Anderson/Fincher/Nolan/Villeneuve/Kubrick), node pipeline architecture, scene generation formulas, and 25+ design styles with exact prompts. Tier-1 generation via Higgsfield CLI (30+ models — Soul V2, Nano Banana 2, Veo 3.1, Kling 3.0, Seedance 2.0, Flux 2, Marketing Studio). Triggers on: 'generate content', 'cinematic', 'create scene', 'start frame', 'style library', 'generate campaign', 'AI filmmaking', 'higgsfield generate', 'soul cinema', 'marketing studio'.

Visual DNA Compiler -- compiles raw brand assets, character references, and style inspiration into persistent, structured identity files. Uses Gemini multimodal analysis to extract color palettes, composition rules, lighting patterns, and texture preferences. Generates tool-specific prompt fragments for Nano Banana Pro, Soul Cinema, Weavy, and ComfyUI. Inspired by Karpathy's LLM Wiki (compile-then-query) and MemPalace spatial hierarchy. Triggers on: 'compile brand', 'extract DNA', 'character sheet', 'style guide', 'compile identity', 'brand analysis'.

Content Loop + Distribution — compounds existing broomva skills into a unified content lifecycle. Orchestrates /blog-post, /content-creation, /social-intelligence, /brainrot-for-good, and /brand-icons for multi-platform distribution. Adds performance tracking, feedback-driven identity refinement, content calendar, and campaign management. Triggers on: 'distribute content', 'content loop', 'publish campaign', 'content calendar', 'track performance'.

OpenCaptions extension for Content Engine — adds intent-driven CWI (Caption With Intent) captions to the post-production pipeline. Hooks into the grade → caption → final stage. Generates captions that understand video intent (pitch, volume, emotion, emphasis) and style themselves accordingly with variable font weight, size, and color. Uses the OpenCaptions CLI or MCP server. Triggers on: 'add captions', 'opencaptions', 'CWI captions', 'intent captions'.

Meta-review skill for validating generated creative assets (videos, images, designs) against a reference style brief. Extracts frames, compares against style criteria, scores adherence, and produces actionable feedback for iteration. Self-improving feedback loop: each review refines the style brief for the next generation. Use when: (1) reviewing a generated video against a reference style, (2) validating visual quality of AI-generated content, (3) scoring style adherence of a Remotion composition, (4) comparing before/after creative iterations, (5) building a self-improving creative pipeline. Triggers on: 'review video', 'check style', 'creative review', 'style adherence', 'compare to reference', 'validate video', 'review reel', 'quality check'.