[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M77WF29MGMHDZ2XRA6SHAE50":3,"skill-same:01M77WF29MGMHDZ2XRA6SHAE50":58},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":24,"commitSha":25,"skillDirectory":26,"forksCount":27,"repositoryHtmlUrl":28,"repositoryHomepageUrl":6,"ownerHtmlUrl":29,"ownerBlogUrl":30,"ownerDescription":6,"updatedAt":31,"install":32,"resources":45},"01M77WF29MGMHDZ2XRA6SHAE50","spark-training-gotchas",null,"Preflight and diagnose the ten known failure modes for ML training on NVIDIA DGX Spark. Use when a training run on DGX Spark fails to start, OOMs below the 128GB limit, slows down mid-run, or before any multi-hour training job on GB10.","Reihantt6\u002Fsns-myagent","Reihantt6","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F219822095?v=4","01M77KK94PV8B7C2D9XWXNS8DJ","Mohamad Reihan tegar prasetya","01M77KK958QFMTHYXGXFV7KZ79",2,"MIT","https:\u002F\u002Fgithub.com\u002FReihantt6\u002Fsns-myagent\u002Fblob\u002Fmain\u002Fagent\u002Fskills\u002Fspark-training-gotchas\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\ndescription: \"Preflight and diagnose the ten known failure modes for ML training on NVIDIA DGX Spark. Use when a training run on DGX Spark fails to start, OOMs below the 128GB limit, slows down mid-run, or before any multi-hour training job on GB10.\"\n---\n# Spark Training Gotchas\n\nDGX Spark's GB10 chip (Grace Blackwell, SM121, 128GB unified\nmemory, aarch64) has ten recurring failure modes across\nlaunch, memory, thermals, bandwidth, and precision. Each is\nnamed G1–G10 so it can be checked by number — the numbering\nis load-bearing for tooling that runs these checks. Read this\nbefore a long run, not after hour six.\n\n## When to Use This Skill\n\n- A training run fails to start, with an import error or a\n  segfault that doesn't point at the real cause.\n- A run OOMs while `nvidia-smi` still shows headroom.\n- Throughput degrades partway through a run that started fine.\n- Before any multi-hour or multi-epoch job on GB10.\n- Wiring two Sparks together, before picking a parallelism\n  strategy.\n- Choosing between FP8 and NVFP4 for a Spark-hosted run.\n\n## Common Issues Quick Reference\n\n| # | Symptom | Fix |\n|---|---|---|\n| G1 | undefined symbol \u002F segfault | cu130 wheel or container |\n| G2 | flash-attn wrong backend used | skip pip build; monkeypatch on NGC |\n| G3 | OOM despite headroom | drop page cache |\n| G4 | throughput drop \u002F reboot | expect ~100W sustained cap |\n| G5 | memory-bound step slow | budget 180–192 GB\u002Fs |\n| G6 | cache evicted mid-run | one GPU server at a time |\n| G7 | NVFP4 slower than FP8 | stay FP8 unless `sm_121a` |\n| G8 | playbook fails outright | check upstream issues |\n| G9 | env breaks after install | use a container |\n| G10 | 2-Spark TP hangs | DDP\u002FFSDP only, never TP |\n\n## The Ten Gotchas\n\n### G1: CUDA 12\u002F13 ABI Mismatch\n\n- **SYMPTOM:** `ImportError: undefined symbol` naming a CUDA\n  function, or a segfault on the first `.cuda()` call.\n- **CAUSE:** most PyPI wheels link `libcudart.so.12`; Spark\n  ships CUDA 13. pip never checks CUDA ABI, so it surfaces\n  only at import or first kernel launch.\n- **CHECK:** `references\u002Fgotcha-checks.md` G1 — the wheel's\n  CUDA build tag.\n- **FIX:** reinstall from `download.pytorch.org\u002Fwhl\u002Fcu130` or\n  use a matched container.\n\n### G2: flash-attn — Skip the pip Build, Watch Unsloth's Auto-Detect\n\n- **SYMPTOM:** `pip install flash-attn` still fails\u002Fhangs.\n  Unsloth may also silently train flash-attn over an\n  explicitly requested SDPA.\n- **CAUSE:** no aarch64\u002Fsm_121 wheel for bare pip — but NGC\n  containers ship a working SM121 flash-attn, and Unsloth\n  auto-prefers it, dropping `attn_implementation=\"sdpa\"`.\n- **CHECK:** `references\u002Fgotcha-checks.md` G2 — is flash-attn\n  already present and working.\n- **FIX:** bare pip — skip flash-attn, use SDPA (unchanged). On\n  NGC — the only reliable override is the monkeypatch in\n  `references\u002Fgotcha-checks.md` G2.\n\n### G3: UMA OOM Below 128GB\n\n- **SYMPTOM:** OOM during model load\u002Ftraining while\n  `nvidia-smi` still reports free memory under the 128GB cap\n  — or, on some setups, `[N\u002FA]` outright instead of a number.\n- **CAUSE:** mmap and the CUDA allocator double-count pages\n  during safetensors load; QLoRA can OOM *earlier* than bf16\n  since dequantization adds transient allocs.\n- **CHECK:** `references\u002Fgotcha-checks.md` G3 — read `free -g`\n  and `\u002Fproc\u002Fmeminfo`, not `nvidia-smi`.\n- **FIX:** drop the page cache with\n  `sync; echo 3 > \u002Fproc\u002Fsys\u002Fvm\u002Fdrop_caches` — needs root, a\n  between-run reset, not a mid-training step.\n\n### G4: Thermal Throttling\n\n- **SYMPTOM:** throughput drops partway through a multi-hour\n  run, or the box spontaneously reboots under sustained load.\n- **CAUSE:** sustained power draw caps around 100W versus the\n  240W rated figure; long runs push into that ceiling and\n  throttle or, sometimes, reboot.\n- **CHECK:** `references\u002Fgotcha-checks.md` G4 — sample\n  `nvidia-smi --query-gpu=temperature.gpu,power.draw`.\n- **FIX:** if power plateaus under 240W while temperature\n  climbs, treat throttling as the cause; improve cooling or\n  cap run length.\n\n### G5: Bandwidth Ceiling\n\n- **SYMPTOM:** memory-bound workloads, decode-heavy RL loops\n  especially, plateau well below expected throughput.\n- **CAUSE:** 273 GB\u002Fs is a spec ceiling, not sustained;\n  measured bandwidth runs 180–192 GB\u002Fs.\n- **CHECK:** `references\u002Fgotcha-checks.md` G5 — observed step\n  time vs. the measured range, not spec.\n- **FIX:** budget throughput from 180–192 GB\u002Fs; revise a plan\n  built on the 273 GB\u002Fs figure.\n\n### G6: Global UMA Resource Contention\n\n- **SYMPTOM:** a process's KV cache\u002Fweights get evicted\n  mid-run silently, no OOM in its own logs.\n- **CAUSE:** unified memory is\n  one global pool; an uncapped\n  or near-capacity process\n  competes with anything else\n  and can evict it. A small,\n  bounded workload doesn't — a\n  \u003C4GB LoRA coexists fine\n  alongside vLLM capped at\n  `gpu-memory-utilization\u003C=0.5`.\n- **CHECK:** `references\u002Fgotcha-checks.md`\n  G6 — other GPU-resident\n  processes and whether\n  capped.\n- **FIX:** the one-heavy-job\n  rule applies to **uncapped or\n  near-capacity** workloads —\n  cap or stop unrelated servers\n  first. A small, capped\n  workload need not\n  stop.\n\n### G7: NVFP4 Slower Than FP8 on SM121\n\n- **SYMPTOM:** switching an inference workload from FP8 to\n  NVFP4 on Spark makes it slower, not faster.\n- **CAUSE:** SM121 lacks `cvt.e2m1x2` unless kernels target\n  `sm_121a`; NVFP4 runs ~32% slower without it.\n- **CHECK:** `references\u002Fgotcha-checks.md` G7 — capability\n  reports `(12, 1)`; does the build target `sm_121a`?\n- **FIX:** stay on FP8 unless the build targets `sm_121a`.\n\n### G8: Stale Official Playbooks\n\n- **SYMPTOM:** following an official DGX Spark playbook still\n  fails, with no local misconfiguration explaining it.\n- **CAUSE:** official playbooks have shipped broken before;\n  the stack moves faster than the docs.\n- **CHECK:** `references\u002Fgotcha-checks.md` G8 — the playbook\n  repo's recent issues.\n- **FIX:** check `github.com\u002FNVIDIA\u002Fdgx-spark-playbooks` issues\n  before trusting a recipe for an expensive run.\n\n### G9: Container-First, Not Bare Pip\n\n- **SYMPTOM:** a bare-pip environment that worked yesterday\n  breaks after an unrelated `pip install`, or two \"identical\"\n  environments behave differently.\n- **CAUSE:** bare pip lets Triton, xformers, and transformers\n  drift independently; nothing pins them to GB10's SM121\n  target.\n- **CHECK:** `references\u002Fgotcha-checks.md`\n  G9 — container or bare pip?\n- **FIX:** prefer an NGC container (see `spark-environment-setup`\n  for tag guidance) or Unsloth's container. If bare pip is\n  unavoidable, follow the NVIDIA install order, including\n  `--no-deps` on Unsloth.\n\n### G10: Dual-Spark Is DDP\u002FFSDP Only\n\n- **SYMPTOM:** a tensor-parallel launch across two Sparks\n  hangs, runs far slower than single-Spark, or errors out.\n- **CAUSE:** ConnectX-7 is fast enough for gradient\u002Fparameter\n  sync (DDP, FSDP) but too thin for TP's fine-grained traffic.\n- **CHECK:** `references\u002Fgotcha-checks.md` G10 — the\n  configured parallelism strategy.\n- **FIX:** on a two-Spark setup, choose DDP or FSDP, never\n  tensor parallelism — TP is single-node only here.\n\n## Fast Triage\n\nThe cheapest checks to run before anything else:\n\n```bash\npython3 -c \"import torch; print(torch.version.cuda)\"  # expect 13.x (G1); NGC builds have no +cu130 tag — that's not a failure\n```\n\n```python\nimport torch; print(torch.cuda.get_device_capability())  # expect (12, 1) (G7)\n```\n\n```bash\n{ [ -f \u002F.dockerenv -o -f \u002Frun\u002F.containerenv ] || grep -qE 'docker|containerd' \u002Fproc\u002F1\u002Fcgroup; } 2>\u002Fdev\u002Fnull && echo container || echo unknown  # G9\n```\n\n`assets\u002Fpreflight.sh` runs G1, G3, G4, G7, G9 and produces one\noutput line per gotcha in a fixed format: G-number first, then\nPASS\u002FFAIL\u002FWARN where automatable, SKIP when unavailable, or\n`INFO:` for a raw reading (G3, G4). Full commands:\n`references\u002Fgotcha-checks.md`. See also\n`spark-environment-setup` for the environment assumed working.\n","# Spark Training Gotchas\n\nDGX Spark's GB10 chip (Grace Blackwell, SM121, 128GB unified\nmemory, aarch64) has ten recurring failure modes across\nlaunch, memory, thermals, bandwidth, and precision. Each is\nnamed G1–G10 so it can be checked by number — the numbering\nis load-bearing for tooling that runs these checks. Read this\nbefore a long run, not after hour six.\n\n## When to Use This Skill\n\n- A training run fails to start, with an import error or a\n  segfault that doesn't point at the real cause.\n- A run OOMs while `nvidia-smi` still shows headroom.\n- Throughput degrades partway through a run that started fine.\n- Before any multi-hour or multi-epoch job on GB10.\n- Wiring two Sparks together, before picking a parallelism\n  strategy.\n- Choosing between FP8 and NVFP4 for a Spark-hosted run.\n\n## Common Issues Quick Reference\n\n| # | Symptom | Fix |\n|---|---|---|\n| G1 | undefined symbol \u002F segfault | cu130 wheel or container |\n| G2 | flash-attn wrong backend used | skip pip build; monkeypatch on NGC |\n| G3 | OOM despite headroom | drop page cache |\n| G4 | throughput drop \u002F reboot | expect ~100W sustained cap |\n| G5 | memory-bound step slow | budget 180–192 GB\u002Fs |\n| G6 | cache evicted mid-run | one GPU server at a time |\n| G7 | NVFP4 slower than FP8 | stay FP8 unless `sm_121a` |\n| G8 | playbook fails outright | check upstream issues |\n| G9 | env breaks after install | use a container |\n| G10 | 2-Spark TP hangs | DDP\u002FFSDP only, never TP |\n\n## The Ten Gotchas\n\n### G1: CUDA 12\u002F13 ABI Mismatch\n\n- **SYMPTOM:** `ImportError: undefined symbol` naming a CUDA\n  function, or a segfault on the first `.cuda()` call.\n- **CAUSE:** most PyPI wheels link `libcudart.so.12`; Spark\n  ships CUDA 13. pip never checks CUDA ABI, so it surfaces\n  only at import or first kernel launch.\n- **CHECK:** `references\u002Fgotcha-checks.md` G1 — the wheel's\n  CUDA build tag.\n- **FIX:** reinstall from `download.pytorch.org\u002Fwhl\u002Fcu130` or\n  use a matched container.\n\n### G2: flash-attn — Skip the pip Build, Watch Unsloth's Auto-Detect\n\n- **SYMPTOM:** `pip install flash-attn` still fails\u002Fhangs.\n  Unsloth may also silently train flash-attn over an\n  explicitly requested SDPA.\n- **CAUSE:** no aarch64\u002Fsm_121 wheel for bare pip — but NGC\n  containers ship a working SM121 flash-attn, and Unsloth\n  auto-prefers it, dropping `attn_implementation=\"sdpa\"`.\n- **CHECK:** `references\u002Fgotcha-checks.md` G2 — is flash-attn\n  already present and working.\n- **FIX:** bare pip — skip flash-attn, use SDPA (unchanged). On\n  NGC — the only reliable override is the monkeypatch in\n  `references\u002Fgotcha-checks.md` G2.\n\n### G3: UMA OOM Below 128GB\n\n- **SYMPTOM:** OOM during model load\u002Ftraining while\n  `nvidia-smi` still reports free memory under the 128GB cap\n  — or, on some setups, `[N\u002FA]` outright instead of a number.\n- **CAUSE:** mmap and the CUDA allocator double-count pages\n  during safetensors load; QLoRA can OOM *earlier* than bf16\n  since dequantization adds transient allocs.\n- **CHECK:** `references\u002Fgotcha-checks.md` G3 — read `free -g`\n  and `\u002Fproc\u002Fmeminfo`, not `nvidia-smi`.\n- **FIX:** drop the page cache with\n  `sync; echo 3 > \u002Fproc\u002Fsys\u002Fvm\u002Fdrop_caches` — needs root, a\n  between-run reset, not a mid-training step.\n\n### G4: Thermal Throttling\n\n- **SYMPTOM:** throughput drops partway through a multi-hour\n  run, or the box spontaneously reboots under sustained load.\n- **CAUSE:** sustained power draw caps around 100W versus the\n  240W rated figure; long runs push into that ceiling and\n  throttle or, sometimes, reboot.\n- **CHECK:** `references\u002Fgotcha-checks.md` G4 — sample\n  `nvidia-smi --query-gpu=temperature.gpu,power.draw`.\n- **FIX:** if power plateaus under 240W while temperature\n  climbs, treat throttling as the cause; improve cooling or\n  cap run length.\n\n### G5: Bandwidth Ceiling\n\n- **SYMPTOM:** memory-bound workloads, decode-heavy RL loops\n  especially, plateau well below expected throughput.\n- **CAUSE:** 273 GB\u002Fs is a spec ceiling, not sustained;\n  measured bandwidth runs 180–192 GB\u002Fs.\n- **CHECK:** `references\u002Fgotcha-checks.md` G5 — observed step\n  time vs. the measured range, not spec.\n- **FIX:** budget throughput from 180–192 GB\u002Fs; revise a plan\n  built on the 273 GB\u002Fs figure.\n\n### G6: Global UMA Resource Contention\n\n- **SYMPTOM:** a process's KV cache\u002Fweights get evicted\n  mid-run silently, no OOM in its own logs.\n- **CAUSE:** unified memory is\n  one global pool; an uncapped\n  or near-capacity process\n  competes with anything else\n  and can evict it. A small,\n  bounded workload doesn't — a\n  \u003C4GB LoRA coexists fine\n  alongside vLLM capped at\n  `gpu-memory-utilization\u003C=0.5`.\n- **CHECK:** `references\u002Fgotcha-checks.md`\n  G6 — other GPU-resident\n  processes and whether\n  capped.\n- **FIX:** the one-heavy-job\n  rule applies to **uncapped or\n  near-capacity** workloads —\n  cap or stop unrelated servers\n  first. A small, capped\n  workload need not\n  stop.\n\n### G7: NVFP4 Slower Than FP8 on SM121\n\n- **SYMPTOM:** switching an inference workload from FP8 to\n  NVFP4 on Spark makes it slower, not faster.\n- **CAUSE:** SM121 lacks `cvt.e2m1x2` unless kernels target\n  `sm_121a`; NVFP4 runs ~32% slower without it.\n- **CHECK:** `references\u002Fgotcha-checks.md` G7 — capability\n  reports `(12, 1)`; does the build target `sm_121a`?\n- **FIX:** stay on FP8 unless the build targets `sm_121a`.\n\n### G8: Stale Official Playbooks\n\n- **SYMPTOM:** following an official DGX Spark playbook still\n  fails, with no local misconfiguration explaining it.\n- **CAUSE:** official playbooks have shipped broken before;\n  the stack moves faster than the docs.\n- **CHECK:** `references\u002Fgotcha-checks.md` G8 — the playbook\n  repo's recent issues.\n- **FIX:** check `github.com\u002FNVIDIA\u002Fdgx-spark-playbooks` issues\n  before trusting a recipe for an expensive run.\n\n### G9: Container-First, Not Bare Pip\n\n- **SYMPTOM:** a bare-pip environment that worked yesterday\n  breaks after an unrelated `pip install`, or two \"identical\"\n  environments behave differently.\n- **CAUSE:** bare pip lets Triton, xformers, and transformers\n  drift independently; nothing pins them to GB10's SM121\n  target.\n- **CHECK:** `references\u002Fgotcha-checks.md`\n  G9 — container or bare pip?\n- **FIX:** prefer an NGC container (see `spark-environment-setup`\n  for tag guidance) or Unsloth's container. If bare pip is\n  unavoidable, follow the NVIDIA install order, including\n  `--no-deps` on Unsloth.\n\n### G10: Dual-Spark Is DDP\u002FFSDP Only\n\n- **SYMPTOM:** a tensor-parallel launch across two Sparks\n  hangs, runs far slower than single-Spark, or errors out.\n- **CAUSE:** ConnectX-7 is fast enough for gradient\u002Fparameter\n  sync (DDP, FSDP) but too thin for TP's fine-grained traffic.\n- **CHECK:** `references\u002Fgotcha-checks.md` G10 — the\n  configured parallelism strategy.\n- **FIX:** on a two-Spark setup, choose DDP or FSDP, never\n  tensor parallelism — TP is single-node only here.\n\n## Fast Triage\n\nThe cheapest checks to run before anything else:\n\n```bash\npython3 -c \"import torch; print(torch.version.cuda)\"  # expect 13.x (G1); NGC builds have no +cu130 tag — that's not a failure\n```\n\n```python\nimport torch; print(torch.cuda.get_device_capability())  # expect (12, 1) (G7)\n```\n\n```bash\n{ [ -f \u002F.dockerenv -o -f \u002Frun\u002F.containerenv ] || grep -qE 'docker|containerd' \u002Fproc\u002F1\u002Fcgroup; } 2>\u002Fdev\u002Fnull && echo container || echo unknown  # G9\n```\n\n`assets\u002Fpreflight.sh` runs G1, G3, G4, G7, G9 and produces one\noutput line per gotcha in a fixed format: G-number first, then\nPASS\u002FFAIL\u002FWARN where automatable, SKIP when unavailable, or\n`INFO:` for a raw reading (G3, G4). Full commands:\n`references\u002Fgotcha-checks.md`. See also\n`spark-environment-setup` for the environment assumed working.\n",{"description":7},"main","58843b159d7a4050d6b785a98b908cce505c0b92","agent\u002Fskills\u002Fspark-training-gotchas",0,"https:\u002F\u002Fgithub.com\u002FReihantt6\u002Fsns-myagent","https:\u002F\u002Fgithub.com\u002FReihantt6","reihsns.web.id","2026-08-19 20:08:18",{"source":8,"skillName":5,"commands":33,"prompt":43,"zipAvailable":44},[34,37,40],{"tool":35,"command":36},"npx","npx skills add Reihantt6\u002Fsns-myagent --skill spark-training-gotchas",{"tool":38,"command":39},"bunx","bunx skills add Reihantt6\u002Fsns-myagent --skill spark-training-gotchas",{"tool":41,"command":42},"pnpm","pnpm dlx skills add Reihantt6\u002Fsns-myagent --skill spark-training-gotchas","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FReihantt6\u002Fsns-myagent\u002Fblob\u002Fmain\u002Fagent\u002Fskills\u002Fspark-training-gotchas\u002FSKILL.md",true,[46,52],{"path":47,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":50,"sourceHtmlUrl":51},"assets\u002Fpreflight.sh","SCRIPT","脚本",2486,"https:\u002F\u002Fgithub.com\u002FReihantt6\u002Fsns-myagent\u002Fblob\u002Fmain\u002Fagent\u002Fskills\u002Fspark-training-gotchas\u002Fassets\u002Fpreflight.sh",{"path":53,"resourceType":54,"resourceTypeDesc":55,"sizeBytes":56,"sourceHtmlUrl":57},"references\u002Fgotcha-checks.md","REFERENCE","参考文档",9598,"https:\u002F\u002Fgithub.com\u002FReihantt6\u002Fsns-myagent\u002Fblob\u002Fmain\u002Fagent\u002Fskills\u002Fspark-training-gotchas\u002Freferences\u002Fgotcha-checks.md",[59,66,73,80,87,94],{"skillNo":60,"name":61,"nameZh":6,"description":62,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":63,"categories":64,"keywords":65,"publishedAt":19},"01M77XDCYZ407XYMF2T4P6F7VT","youtube-transcript","Extract transcripts from YouTube videos. Use when the user asks for a transcript, subtitles, or captions of a YouTube video and provides a YouTube URL (youtube.com\u002Fwatch?v=, youtu.be\u002F, or similar). Supports output with or without timestamps.","https:\u002F\u002Fgithub.com\u002FReihantt6\u002Fsns-myagent\u002Fblob\u002Fmain\u002Fagent\u002Fskills\u002Fyoutube-transcript\u002FSKILL.md",[],[],{"skillNo":67,"name":68,"nameZh":6,"description":69,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":70,"categories":71,"keywords":72,"publishedAt":19},"01M77XD7PBC98J6MXGJJ2TJ8DX","youtube-thumbnail-design","YouTube thumbnail design with specific dimensions, contrast rules, and mobile preview optimization. Covers safe zones, text placement, face expression psychology, and A\u002FB testing. Use for: YouTube thumbnails, video cover images, click-through optimization. Triggers: youtube thumbnail, thumbnail design, video thumbnail, click through rate, ctr optimization, youtube cover, video cover image, thumbnail maker, thumbnail tips, youtube design, video preview image","https:\u002F\u002Fgithub.com\u002FReihantt6\u002Fsns-myagent\u002Fblob\u002Fmain\u002Fagent\u002Fskills\u002Fyoutube-thumbnail-design\u002FSKILL.md",[],[],{"skillNo":74,"name":75,"nameZh":6,"description":76,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":77,"categories":78,"keywords":79,"publishedAt":19},"01M77XD2541C8HXJ6RHYZP4NGJ","youtube-short-maker","Make YouTube Shorts that hold attention with Pexo — vertical 9:16, fast-paced, hook in the first 3 seconds. Describe your topic and Pexo writes the script, generates the shots, picks the models, and assembles a finished Short with music and captions. Use for YouTube Shorts: \"youtube short\", \"make a youtube short\", \"shorts video\", \"short-form video\". NOT for long-form or landscape video.","https:\u002F\u002Fgithub.com\u002FReihantt6\u002Fsns-myagent\u002Fblob\u002Fmain\u002Fagent\u002Fskills\u002Fyoutube-short-maker\u002FSKILL.md",[],[],{"skillNo":81,"name":82,"nameZh":6,"description":83,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":84,"categories":85,"keywords":86,"publishedAt":19},"01M77XCXKVSYK7KTZ3PQBVVGFD","xcode-project-setup","Safely modifies Xcode projects (.pbxproj) to add Swift Packages and link files. Use this skill whenever an iOS project needs dependencies installed (e.g. Firebase, Alamofire).","https:\u002F\u002Fgithub.com\u002FReihantt6\u002Fsns-myagent\u002Fblob\u002Fmain\u002Fagent\u002Fskills\u002Fxcode-project-setup\u002FSKILL.md",[],[],{"skillNo":88,"name":89,"nameZh":6,"description":90,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":91,"categories":92,"keywords":93,"publishedAt":19},"01M77XCR1DV8ZQK2NHWKVY4EWX","x402","Set up Browser Use Cloud payments with x402 — pay per request from a crypto wallet (USDC on Base mainnet), no signup or API key. Two setups it works out up front — \"just use it\" (set up a wallet so you or Claude Code can run cloud browser tasks paid from the wallet — Claude writes and runs throwaway scripts, nothing touches your codebase) or \"build it in\" (install the SDK and write the key + code into your project). Walks through wallet setup, funding, .env, and a ~$1 test run. Use when the user asks about x402, pay-per-use, USDC payments, or wants Browser Use Cloud without an API key. For the free-tier signup (reverse-CAPTCHA → API key), use `browser-use cloud signup` or the `cloud` skill instead.","https:\u002F\u002Fgithub.com\u002FReihantt6\u002Fsns-myagent\u002Fblob\u002Fmain\u002Fagent\u002Fskills\u002Fx402\u002FSKILL.md",[],[],{"skillNo":95,"name":96,"nameZh":6,"description":97,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":98,"categories":99,"keywords":100,"publishedAt":19},"01M77XCMD3R3ZBZQQSW6E0793B","x-twitter-scraper","Build GitHub Copilot workflows with Xquik X API SDKs, REST endpoints, hosted Apify Actor runs, MCP tools, TweetClaw OpenClaw plugin installs, signed webhooks, tweet search, user lookup, follower exports, media actions, and agent automation.","https:\u002F\u002Fgithub.com\u002FReihantt6\u002Fsns-myagent\u002Fblob\u002Fmain\u002Fagent\u002Fskills\u002Fx-twitter-scraper\u002FSKILL.md",[],[]]