[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7F1A2D2ZASF4FR1HB9V2E1M":3,"skill-same:01M7F1A2D2ZASF4FR1HB9V2E1M":56},{"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":14,"repositoryHtmlUrl":27,"repositoryHomepageUrl":6,"ownerHtmlUrl":28,"ownerBlogUrl":29,"ownerDescription":30,"updatedAt":31,"install":32,"resources":45},"01M7F1A2D2ZASF4FR1HB9V2E1M","dataset-curation",null,"Prepare, format, and validate datasets for supervised fine-tuning and preference training. Use when converting raw data into training format, applying chat templates, configuring sequence packing, generating synthetic training data, or writing a dataset card before a run.","shotcowboystyle\u002Fag-skills","shotcowboystyle","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F62471?v=4","01M7EZTCFCZJ6K3MYW8BPM4X6R","Curtis Blanton","01M7EZTCG1S5M04J90A931X9BX",0,"MIT","https:\u002F\u002Fgithub.com\u002Fshotcowboystyle\u002Fag-skills\u002Fblob\u002Fmain\u002Fplugins\u002Fllm-finetuning\u002Fskills\u002Fdataset-curation\u002FSKILL.md",[],[],"2026-11-16 10:05:34",false,"---\nname: dataset-curation\ndescription: Prepare, format, and validate datasets for supervised fine-tuning and preference training. Use when converting raw data into training format, applying chat templates, configuring sequence packing, generating synthetic training data, or writing a dataset card before a run.\n---\n\n# Dataset Curation\n\nThis skill assumes `finetuning-method-selection`\nalready routed here — the next step is preparing\ndata, not choosing a method. What follows: format\nselection by target method, the template\u002Fpacking\nmechanics behind the most common silent training\nfailures, rules for mixing in synthetic data\nwithout collapse, and the dataset card that closes\nout Phase 2 before a run starts.\n\n**Input:** raw examples (demonstrations, preference\njudgments, or task prompts) plus a routing decision\nfrom `finetuning-method-selection`.\n**Output format:** a formatted, packed, validated\nJSONL dataset plus a completed dataset card — the\nPhase 2 artifact `\u002Ffinetune` checks before launching\ntraining.\n\n## Format Selection\n\n| Method | Shape | Rows |\n|---|---|---|\n| SFT, single-turn | Instruct (`instruction`\u002F`response` or `prompt`\u002F`completion`) | ~1,000+ floor |\n| SFT, multi-turn | Conversation \u002F ChatML `messages` list | ~1,000+ floor |\n| DPO \u002F ORPO | Preference pair (`prompt`, `chosen`, `rejected`) | Method-dependent, see `preference-optimization` |\n| KTO | Unpaired (`prompt`, `completion`, `label`) | Method-dependent, see `preference-optimization` |\n| GRPO \u002F RLVR | Prompt-only (`prompt` + verifier metadata) | Method-dependent, see `grpo-rlvr-training` |\n\n- **~1,000+ rows is the recommended floor for SFT**,\n  not a target. Below it, a handful of low-quality\n  or duplicate examples can dominate the gradient;\n  above it, **quality over quantity** — a smaller\n  verified, deduplicated set beats a larger noisy one.\n- The ChatML shape, for orientation; the other four\n  formats plus a ShareGPT conversion note live in\n  `references\u002Fformats-and-templates.md`:\n\n  ```json\n  {\"messages\": [\n    {\"role\": \"user\", \"content\": \"...\"},\n    {\"role\": \"assistant\", \"content\": \"...\"}\n  ]}\n  ```\n\n## Chat Templates and Loss Masking\n\nApply the target model's chat template **before**\nany concatenation or packing, never after — packing\nraw text and templating the packed blob afterward\ncorrupts turn boundaries, landing role markers in\nthe wrong place relative to each example.\n\n- **Train on assistant responses only.** Mask the\n  loss (`-100` in the labels tensor) over system\u002Fuser\n  turns and the template's own role markers — only\n  assistant-turn content tokens contribute to loss.\n- **Template\u002Ftokenizer mismatches are a top silent\n  failure mode.** A model trained against one chat\n  template but served or evaluated with a different\n  one degrades without erroring. Verify the same\n  template string used in training is applied at\n  inference and eval time.\n- **Keep the dataset in `messages` shape** and let\n  the trainer template and mask it\n  (`assistant_only_loss=True` in current TRL) —\n  pre-rendering to a flat text field destroys the\n  turn boundaries masking needs. Full code sketch:\n  `references\u002Fformats-and-templates.md`. Sanity-check\n  before training — decode only unmasked positions;\n  expect only assistant text:\n\n  ```python\n  keep = batch[\"labels\"][0] != -100\n  print(tokenizer.decode(batch[\"input_ids\"][0][keep]))\n  ```\n\n## Packing\n\n**Without packing, compute goes to padding** —\nvariable-length examples batched at a fixed\nsequence length waste the gap between each\nexample's length and the batch's max. How much\nneeds measurement: compare the sum of example\ntoken counts with `batches × max_length` on your\ndata. Packing concatenates multiple examples into\none sequence up to the max length, cutting most of\nthat waste. Padding-free (position-id) packing\nkeeps examples from attending to each other\nwithout a block-diagonal mask; prefer it where the\ntrainer supports it.\n\n- **Packing changes batch semantics.** A packed\n  sequence can contain several original examples, so\n  \"steps per epoch\" and any LR schedule keyed to\n  example count shift once packing is on — recompute\n  schedule milestones against packed-sequence count.\n- **MANDATORY: decode and manually inspect 5–10\n  packed sequences before scaling to a full run.**\n  Confirm example boundaries land where expected,\n  template markers are intact per sub-example, and\n  the loss mask is still assistant-only within each\n  packed sequence. Not optional — packing bugs are\n  silent (the loss curve looks normal) and only\n  surface in eval quality, hours later:\n\n  ```python\n  for seq in packed_dataset.select(range(10)):\n      print(tokenizer.decode(seq[\"input_ids\"]))\n  ```\n\n## Synthetic Data Rules\n\n- **Keep ≥25% real data as a collapse guard.**\n  Training on a growing share of model-generated\n  data without a real-data floor drives measurable\n  quality collapse over successive generations —\n  25% real is the minimum that holds the line.\n  **General-domain replay rows\n  count toward this floor** —\n  \"real\" means \"not generated\n  for this task from this\n  student,\" not \"human-authored.\"\n  An all-synthetic-by-construction\n  dataset can meet the ≥25% floor\n  through replay alone (see\n  `references\u002Fsynthetic-data.md`'s\n  Replay-Mix Construction recipe);\n  state which rows count as \"real\"\n  in the dataset card rather than\n  leaving the floor structurally\n  unmeetable.\n- **Magpie and rejection sampling are the\n  workhorses.** Magpie extracts prompts from the\n  model's own template prior; rejection sampling\n  generates several candidates per prompt and keeps\n  only the ones a filter passes. Both beat naive\n  single-shot generation.\n- **Targeted, student-aware generation usually\n  beats static generation** — aiming at the\n  student's actual failure modes hits a quality bar\n  with fewer filtered examples. The gain needs\n  measurement on your task.\n- **Measure the accept rate on a pilot batch** and\n  plan volume from it: raw generations needed ≈\n  target rows ÷ accept rate.\n- Generation-method ranking, filter funnel, replay-\n  mix construction, and distillation pattern:\n  `references\u002Fsynthetic-data.md`.\n\n## The Dataset Card\n\nEvery dataset that reaches training gets a card —\nthe required Phase 2 artifact `\u002Ffinetune` checks\nbefore launching. The card is not free-form\ndocumentation; it MUST carry these fields:\n\n- **Provenance** — where every row came from (real\n  source(s), synthetic method(s), or both),\n  traceable to `trace-to-training-data` output.\n- **Counts** — total rows, and rows per split\n  (train\u002Feval\u002Fheld-out) if split.\n- **Synthetic\u002Freal ratio** — the measured ratio,\n  checked against the ≥25% real floor above.\n- **Dedup method** — exact-match, semantic\n  (embedding threshold), or both; see the filter\n  funnel in `references\u002Fsynthetic-data.md`.\n- **Template used** — the exact chat template\n  string\u002Fidentifier, kept consistent through\n  inference and eval — this is what ties an\n  `eval-harness-first` run back to the checkpoint.\n- **Packing config** — whether packing was used,\n  max sequence length, and confirmation the\n  5–10-sequence manual inspection above was done.\n\nA dataset missing any of these six fields isn't\nready for `\u002Ffinetune` — the card is a gate, not a\nsummary written after the fact.\n\n### Phase 2 Exit Checklist\n\nBefore handing off to `\u002Ffinetune`, confirm:\n\n1. Format matches the method (table above).\n2. Template applied before concatenation.\n3. Loss masked to assistant turns only.\n4. 5–10 packed sequences decoded and read.\n5. ≥25% real data in the final mix.\n6. Dataset card complete — all six fields.\n\n## References\n\n- `references\u002Fformats-and-templates.md` — JSONL\n  examples per format, current-TRL masking code,\n  and the ShareGPT conversion note.\n- `references\u002Fsynthetic-data.md` — generation-method\n  ranking, filter funnel, replay-mix construction,\n  and teacher→student distillation pattern.\n\nRelated skills: `finetuning-method-selection` routes\nhere; `lora-qlora-recipes`, `vision-sft`, and\n`preference-optimization` consume the datasets this\nskill produces; `trace-to-training-data` is the\nprovenance source for graded-trajectory datasets;\n`eval-harness-first` grades the resulting checkpoint.\n","# Dataset Curation\n\nThis skill assumes `finetuning-method-selection`\nalready routed here — the next step is preparing\ndata, not choosing a method. What follows: format\nselection by target method, the template\u002Fpacking\nmechanics behind the most common silent training\nfailures, rules for mixing in synthetic data\nwithout collapse, and the dataset card that closes\nout Phase 2 before a run starts.\n\n**Input:** raw examples (demonstrations, preference\njudgments, or task prompts) plus a routing decision\nfrom `finetuning-method-selection`.\n**Output format:** a formatted, packed, validated\nJSONL dataset plus a completed dataset card — the\nPhase 2 artifact `\u002Ffinetune` checks before launching\ntraining.\n\n## Format Selection\n\n| Method | Shape | Rows |\n|---|---|---|\n| SFT, single-turn | Instruct (`instruction`\u002F`response` or `prompt`\u002F`completion`) | ~1,000+ floor |\n| SFT, multi-turn | Conversation \u002F ChatML `messages` list | ~1,000+ floor |\n| DPO \u002F ORPO | Preference pair (`prompt`, `chosen`, `rejected`) | Method-dependent, see `preference-optimization` |\n| KTO | Unpaired (`prompt`, `completion`, `label`) | Method-dependent, see `preference-optimization` |\n| GRPO \u002F RLVR | Prompt-only (`prompt` + verifier metadata) | Method-dependent, see `grpo-rlvr-training` |\n\n- **~1,000+ rows is the recommended floor for SFT**,\n  not a target. Below it, a handful of low-quality\n  or duplicate examples can dominate the gradient;\n  above it, **quality over quantity** — a smaller\n  verified, deduplicated set beats a larger noisy one.\n- The ChatML shape, for orientation; the other four\n  formats plus a ShareGPT conversion note live in\n  `references\u002Fformats-and-templates.md`:\n\n  ```json\n  {\"messages\": [\n    {\"role\": \"user\", \"content\": \"...\"},\n    {\"role\": \"assistant\", \"content\": \"...\"}\n  ]}\n  ```\n\n## Chat Templates and Loss Masking\n\nApply the target model's chat template **before**\nany concatenation or packing, never after — packing\nraw text and templating the packed blob afterward\ncorrupts turn boundaries, landing role markers in\nthe wrong place relative to each example.\n\n- **Train on assistant responses only.** Mask the\n  loss (`-100` in the labels tensor) over system\u002Fuser\n  turns and the template's own role markers — only\n  assistant-turn content tokens contribute to loss.\n- **Template\u002Ftokenizer mismatches are a top silent\n  failure mode.** A model trained against one chat\n  template but served or evaluated with a different\n  one degrades without erroring. Verify the same\n  template string used in training is applied at\n  inference and eval time.\n- **Keep the dataset in `messages` shape** and let\n  the trainer template and mask it\n  (`assistant_only_loss=True` in current TRL) —\n  pre-rendering to a flat text field destroys the\n  turn boundaries masking needs. Full code sketch:\n  `references\u002Fformats-and-templates.md`. Sanity-check\n  before training — decode only unmasked positions;\n  expect only assistant text:\n\n  ```python\n  keep = batch[\"labels\"][0] != -100\n  print(tokenizer.decode(batch[\"input_ids\"][0][keep]))\n  ```\n\n## Packing\n\n**Without packing, compute goes to padding** —\nvariable-length examples batched at a fixed\nsequence length waste the gap between each\nexample's length and the batch's max. How much\nneeds measurement: compare the sum of example\ntoken counts with `batches × max_length` on your\ndata. Packing concatenates multiple examples into\none sequence up to the max length, cutting most of\nthat waste. Padding-free (position-id) packing\nkeeps examples from attending to each other\nwithout a block-diagonal mask; prefer it where the\ntrainer supports it.\n\n- **Packing changes batch semantics.** A packed\n  sequence can contain several original examples, so\n  \"steps per epoch\" and any LR schedule keyed to\n  example count shift once packing is on — recompute\n  schedule milestones against packed-sequence count.\n- **MANDATORY: decode and manually inspect 5–10\n  packed sequences before scaling to a full run.**\n  Confirm example boundaries land where expected,\n  template markers are intact per sub-example, and\n  the loss mask is still assistant-only within each\n  packed sequence. Not optional — packing bugs are\n  silent (the loss curve looks normal) and only\n  surface in eval quality, hours later:\n\n  ```python\n  for seq in packed_dataset.select(range(10)):\n      print(tokenizer.decode(seq[\"input_ids\"]))\n  ```\n\n## Synthetic Data Rules\n\n- **Keep ≥25% real data as a collapse guard.**\n  Training on a growing share of model-generated\n  data without a real-data floor drives measurable\n  quality collapse over successive generations —\n  25% real is the minimum that holds the line.\n  **General-domain replay rows\n  count toward this floor** —\n  \"real\" means \"not generated\n  for this task from this\n  student,\" not \"human-authored.\"\n  An all-synthetic-by-construction\n  dataset can meet the ≥25% floor\n  through replay alone (see\n  `references\u002Fsynthetic-data.md`'s\n  Replay-Mix Construction recipe);\n  state which rows count as \"real\"\n  in the dataset card rather than\n  leaving the floor structurally\n  unmeetable.\n- **Magpie and rejection sampling are the\n  workhorses.** Magpie extracts prompts from the\n  model's own template prior; rejection sampling\n  generates several candidates per prompt and keeps\n  only the ones a filter passes. Both beat naive\n  single-shot generation.\n- **Targeted, student-aware generation usually\n  beats static generation** — aiming at the\n  student's actual failure modes hits a quality bar\n  with fewer filtered examples. The gain needs\n  measurement on your task.\n- **Measure the accept rate on a pilot batch** and\n  plan volume from it: raw generations needed ≈\n  target rows ÷ accept rate.\n- Generation-method ranking, filter funnel, replay-\n  mix construction, and distillation pattern:\n  `references\u002Fsynthetic-data.md`.\n\n## The Dataset Card\n\nEvery dataset that reaches training gets a card —\nthe required Phase 2 artifact `\u002Ffinetune` checks\nbefore launching. The card is not free-form\ndocumentation; it MUST carry these fields:\n\n- **Provenance** — where every row came from (real\n  source(s), synthetic method(s), or both),\n  traceable to `trace-to-training-data` output.\n- **Counts** — total rows, and rows per split\n  (train\u002Feval\u002Fheld-out) if split.\n- **Synthetic\u002Freal ratio** — the measured ratio,\n  checked against the ≥25% real floor above.\n- **Dedup method** — exact-match, semantic\n  (embedding threshold), or both; see the filter\n  funnel in `references\u002Fsynthetic-data.md`.\n- **Template used** — the exact chat template\n  string\u002Fidentifier, kept consistent through\n  inference and eval — this is what ties an\n  `eval-harness-first` run back to the checkpoint.\n- **Packing config** — whether packing was used,\n  max sequence length, and confirmation the\n  5–10-sequence manual inspection above was done.\n\nA dataset missing any of these six fields isn't\nready for `\u002Ffinetune` — the card is a gate, not a\nsummary written after the fact.\n\n### Phase 2 Exit Checklist\n\nBefore handing off to `\u002Ffinetune`, confirm:\n\n1. Format matches the method (table above).\n2. Template applied before concatenation.\n3. Loss masked to assistant turns only.\n4. 5–10 packed sequences decoded and read.\n5. ≥25% real data in the final mix.\n6. Dataset card complete — all six fields.\n\n## References\n\n- `references\u002Fformats-and-templates.md` — JSONL\n  examples per format, current-TRL masking code,\n  and the ShareGPT conversion note.\n- `references\u002Fsynthetic-data.md` — generation-method\n  ranking, filter funnel, replay-mix construction,\n  and teacher→student distillation pattern.\n\nRelated skills: `finetuning-method-selection` routes\nhere; `lora-qlora-recipes`, `vision-sft`, and\n`preference-optimization` consume the datasets this\nskill produces; `trace-to-training-data` is the\nprovenance source for graded-trajectory datasets;\n`eval-harness-first` grades the resulting checkpoint.\n",{"name":5,"description":7},"main","f846d9a97082ff30d613b55a085cef425d822473","plugins\u002Fllm-finetuning\u002Fskills\u002Fdataset-curation","https:\u002F\u002Fgithub.com\u002Fshotcowboystyle\u002Fag-skills","https:\u002F\u002Fgithub.com\u002Fshotcowboystyle","https:\u002F\u002Fshotcowboystyle.github.io","Building smarter with AI, automation & modern tools. Exploring how far one person can go without a big team using today’s AI.","2026-11-10 04:58:27",{"source":8,"skillName":5,"commands":33,"prompt":43,"zipAvailable":44},[34,37,40],{"tool":35,"command":36},"npx","npx skills add shotcowboystyle\u002Fag-skills --skill dataset-curation",{"tool":38,"command":39},"bunx","bunx skills add shotcowboystyle\u002Fag-skills --skill dataset-curation",{"tool":41,"command":42},"pnpm","pnpm dlx skills add shotcowboystyle\u002Fag-skills --skill dataset-curation","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fshotcowboystyle\u002Fag-skills\u002Fblob\u002Fmain\u002Fplugins\u002Fllm-finetuning\u002Fskills\u002Fdataset-curation\u002FSKILL.md",true,[46,52],{"path":47,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":50,"sourceHtmlUrl":51},"references\u002Fformats-and-templates.md","REFERENCE","参考文档",6610,"https:\u002F\u002Fgithub.com\u002Fshotcowboystyle\u002Fag-skills\u002Fblob\u002Fmain\u002Fplugins\u002Fllm-finetuning\u002Fskills\u002Fdataset-curation\u002Freferences\u002Fformats-and-templates.md",{"path":53,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":54,"sourceHtmlUrl":55},"references\u002Fsynthetic-data.md",12887,"https:\u002F\u002Fgithub.com\u002Fshotcowboystyle\u002Fag-skills\u002Fblob\u002Fmain\u002Fplugins\u002Fllm-finetuning\u002Fskills\u002Fdataset-curation\u002Freferences\u002Fsynthetic-data.md",[57,63,70,77,84,91],{"skillNo":58,"name":59,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":60,"categories":61,"keywords":62,"publishedAt":19},"01M7F1YRY5S5QRJ1V0BJS130QX","no-description","https:\u002F\u002Fgithub.com\u002Fshotcowboystyle\u002Fag-skills\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fno-description\u002FSKILL.md",[],[],{"skillNo":64,"name":65,"nameZh":6,"description":66,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":67,"categories":68,"keywords":69,"publishedAt":19},"01M7F1Y9BV2FJNQJ3NB32JS5CM","good-skill","Fixture skill that satisfies every rule.","https:\u002F\u002Fgithub.com\u002Fshotcowboystyle\u002Fag-skills\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fgood-skill\u002FSKILL.md",[],[],{"skillNo":71,"name":72,"nameZh":6,"description":73,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":74,"categories":75,"keywords":76,"publishedAt":19},"01M7F1XXBQ3NQRXZSS0KAGAR9V","broken-link","Links to a reference file that does not exist.","https:\u002F\u002Fgithub.com\u002Fshotcowboystyle\u002Fag-skills\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fbroken-link\u002FSKILL.md",[],[],{"skillNo":78,"name":79,"nameZh":6,"description":80,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":81,"categories":82,"keywords":83,"publishedAt":19},"01M7F1XBX4R46NR9W0HXQ7XCDX","Bad_Name","Name violates the kebab-case rule and does not match the directory.","https:\u002F\u002Fgithub.com\u002Fshotcowboystyle\u002Fag-skills\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fbad-name\u002FSKILL.md",[],[],{"skillNo":85,"name":86,"nameZh":6,"description":87,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":88,"categories":89,"keywords":90,"publishedAt":19},"01M7F1WNAF8WNEQ7S5HKT5VK0D","skill-authoring","Write, review, and validate portable agent skills in SKILL.md format. Use when creating a new skill, editing an existing one, or asked whether a skill follows the Agent Skills spec.","https:\u002F\u002Fgithub.com\u002Fshotcowboystyle\u002Fag-skills\u002Fblob\u002Fmain\u002Fskills\u002Fskill-authoring\u002FSKILL.md",[],[],{"skillNo":92,"name":93,"nameZh":6,"description":94,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":95,"categories":96,"keywords":97,"publishedAt":19},"01M7F1W30TC0XFNSF7YBRGETK2","weigh-bags","Record scale readings for each bag — empty (tare), outbound, current, final — and reconcile them against the itemised inventory to find unaccounted weight. Use when the user gives a bag weight, says how heavy something is, weighs their luggage, or asks how close to the limit they are. Appends to bags.yaml and runs scripts\u002Freconcile.py.","https:\u002F\u002Fgithub.com\u002Fshotcowboystyle\u002Fag-skills\u002Fblob\u002Fmain\u002Fplugins\u002Ftravel-packing-assistant\u002Fskills\u002Fweigh-bags\u002FSKILL.md",[],[]]