Skills from real work, not evals.

Use when checking resource usage and stopping forgotten background model jobs, servers, or agent-launched processes. Inventory first, stop only clearly in-scope workloads, then verify listeners and processes are gone.

Control a running Blender instance from Hermes via the Blender MCP addon/socket. Use for Blender scene creation, 3D modeling, materials, lighting, animation, viewport screenshots, renders, and bpy Python automation.

Run parallel research agents feeding one cited synthesis: independent lanes, worker contracts, citation-ledger isolation, claim verification, and provenance-stable drafting.

Produce recurring-character multi-world generative video: locked identity plates, portal transitions, keyframe pinning, seam QA, and surgical clip repair.

Produce branded short-form portrait video in HyperFrames: series format contracts, word-timed highlight captions, exact visualization-to-source handoffs, music ducking, and final-artifact verification.

Build, revise, review, and verify HyperFrames HTML/GSAP video projects: deterministic motion-graphics discipline, production quality bar, revision-to-acceptance-criteria workflow, and honest verification.

Install, validate, and operate local image/video generation stacks such as ComfyUI on NVIDIA DGX Spark / GB10 (aarch64, unified memory), including model-fit assessment and ARM-specific troubleshooting.

Build, schedule, and verify Hermes Agent cron automations with pre-run scripts, wakeAgent gates, no-agent watchdogs, and chained handoffs.

Operate and troubleshoot Hermes multi-profile workflows: Kanban boards, cron jobs, gateway runtime, profile env/toolsets, and paper-only demo loops.

Author, wire, and verify native Hermes Agent tools in the source tree: tool module, registry schema, toolsets.py exposure, credentials, restart, and validation.

Use when operating Hermes Kanban systems as either orchestrator or worker: decomposition, lane discipline, worker lifecycle, and handoff quality.

Run, serve, and evaluate LLMs: local GGUF inference via llama.cpp, high-throughput serving via vLLM, and benchmark evaluation via lm-eval-harness.

Train and fine-tune LLMs: LoRA/QLoRA via PEFT and Unsloth, RLHF/DPO/GRPO via TRL and Axolotl, distributed training via PyTorch FSDP. Covers parameter-efficient methods, reward modeling, and distributed scaling.

Build local-first Markdown knowledge-base browsers and read-only MCP servers.

Inspect, build, reset, and verify Unreal Engine projects through the Unreal MCP server, including asset-registry discovery, batch scene automation, safe destructive edits, viewport capture, and local Unreal knowledge-base research.

Build scroll-driven cinematic web films (Next.js/GSAP/Lenis) that pass hard gates: seek-safe determinism, ?t= addressability, frame-state snapshots, zero CLS, honest frame-time budgets, clean console.

Build and verify slide decks: python-pptx generator scripts for 16:9 .pptx, self-contained HTML deck conversion, shared design tokens, live-demo companion editing, and read-back verification.

Operate and recover a headless NVIDIA DGX Spark: evidence-first network re-discovery after DHCP/power events, no-input login recovery, manual-start service management, and SSH tunnel health discipline.

Use when wiring local OpenAI LLMs into Hermes profiles.

Build deterministic paper-trading engines, backtests, and replay harnesses: lookahead-free event ordering, crash-recovery equivalence, unified cost models, and hard verification gates.

Use when delegating coding work to external agent CLIs such as Claude Code, Codex, or OpenCode, and coordinating their outputs safely.

Use when auditing a custom local-LLM benchmark or evaluation harness for bugs, weak discrimination, grader flaws, artifact integrity, reproducibility, privacy, and misleading reporting. Performs a read-only forensic audit of design, code, tasks, oracles, tests, mock runs, and recorded model artifacts; quantifies real impact before ranking findings and fixes.

Build reproducible pipelines that repurpose owned or licensed long-form videos, social videos, podcasts, and articles into short-form vertical videos for Shorts, TikTok, Instagram Reels, and Facebook Reels.

Build deterministic game simulations — time-loop/ghost-replay puzzles, replay-based games, racing packs — with computed (not authored) solutions and headless verification gates that prove determinism.

Build and verify Three.js/WebGL work — single-file CDN artifacts, Vite apps, procedural worlds and games, and shot-based cinematic films — with Node-harness and headless-Chrome/CDP verification that proves the scene actually renders.

Build and verify engine-free raw-WebGL2 browser games — headless completion gates, CPU-side render-pipeline proof, and the WebGL bug signatures that make garbage frames diagnosable without a GPU.

Generate every visual and audio asset for an engine-free WebGL2 browser game procedurally — dual sRGB/linear texture atlases, original car liveries, Canvas 2D HUD, Web Audio synthesis, and headless skia-canvas previews with vision verification.

Independent, evidence-gated QA for local playable browser games — stack/headless/visual/perf gate families, a Playwright harness, deterministic-sim rebaselining, and side-by-side reference-still acceptance. Every PASS points at a file.

Use when a Windows system drive is low on space — evidence-based, category-scoped cleanup of C: including developer caches, installers, app recovery data, and WSL2/Docker VHDX reclaim, with live free-space verification after every phase.

Use when you have a Uniswap v4 pool ID and need its hook, fee, tick spacing and currencies — recover and prove the PoolKey over raw JSON-RPC without an explorer.

Use when building or testing tools that read EVM liquidity data over RPC — curve walking, fee math, route depth, oracle freshness — so missing reads never masquerade as measured zeros.

Use when exhaustively mining an account's or organization's whole GitHub/Hugging Face ecosystem — inventory-first coverage, per-artifact dispositions, quota-free retrieval, and evidence-graded synthesis by technique.

Use when measuring local LLM speed on Ollama, vLLM or llama.cpp: cold start, prefill vs decode tok/s, placement, context sweeps, concurrency, quant speed, unified-memory safety.

Use when designing, building, running, or reviewing an evaluation suite for local/open-weight LLMs: layered scoring, fail-closed readiness, safe execution, judging, and honest reports.

Use when quantizing a local LLM (NVFP4, GGUF, mixed low-bit MoE PTQ) and proving it: scoped comparison contract, coverage probe, host safety, artifact and serving gates.

Use when gating generated video clips before delivery or master assembly: identity, scale, cast, prop, action, dialogue and transition QA, plus clean-endpoint and partial-take recovery.

Use when one character and world must stay identical across generated stills, turnarounds and clips (Krea 2 Turbo img2img/Style Reference, MiniMax H3 I2V/Ref2VA) and stitched episodes.

Use when turning a verbal brand direction into logo and visual-identity concepts, refining a chosen route, or packaging approved art into production vectors or style references.

Use when producing YouTube channel deliverables — thumbnails, transcript-based descriptions, technical video scripts — under a channel's own house format, or identifying a video from its ID.

Use when building or repairing self-contained HTML (decks, prototypes, reports, demos) that must work from file://, desktop previews or embedded viewers that block scripts or storage.

Use when fine-tuning Laya or any typed-decision (System 1) model on your own logged decisions — tool calls, skill loads, approvals — and you need to know whether it actually helped rather than overfit.