[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7CZGYTNQJG0BE9CM7K7XCSW":3,"skill-same:01M7CZGYTNQJG0BE9CM7K7XCSW":60},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":9,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":15,"categories":16,"keywords":17,"publishedAt":18,"linkOnly":19,"markdownContent":20,"markdownBody":21,"frontMatter":22,"refName":32,"commitSha":33,"skillDirectory":34,"forksCount":35,"repositoryHtmlUrl":30,"repositoryHomepageUrl":6,"ownerHtmlUrl":36,"ownerBlogUrl":37,"ownerDescription":38,"updatedAt":39,"install":40,"resources":53},"01M7CZGYTNQJG0BE9CM7K7XCSW","padosoft-rag-ingestion-security",null,"Use this skill when content is brought into a corpus that a model will read — a connector to a document store, a mailbox, a ticket system, a wiki, an upload, a crawl — and whenever the question is who may see what came in. Also when the user reports that an assistant surfaced a document somebody should not have seen, asks how to mirror the source's permissions, worries about content written by outsiders reaching a tool call, or needs personal data kept out of an index. It covers the ingestion contract, permission mirroring that fails closed, authorship provenance, the indirect-injection chain, and redaction before embedding. Do not use it for choosing an embedding model, for retrieval quality, or for the shape of the corpus itself (padosoft-rag-knowledge-base).","padosoft\u002Fskills","padosoft","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F10467753?v=4","01M7CZ3FRSM851DK225E3KK2VF","01M7CZ3FSY016E21VS6EABM21V",1,"MIT","https:\u002F\u002Fgithub.com\u002Fpadosoft\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002Fpadosoft-rag-ingestion-security\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: padosoft-rag-ingestion-security\ndescription: >-\n  Use this skill when content is brought into a corpus that a model will read — a connector to a document\n  store, a mailbox, a ticket system, a wiki, an upload, a crawl — and whenever the question is who may see\n  what came in. Also when the user reports that an assistant surfaced a document somebody should not have\n  seen, asks how to mirror the source's permissions, worries about content written by outsiders reaching a\n  tool call, or needs personal data kept out of an index. It covers the ingestion contract, permission\n  mirroring that fails closed, authorship provenance, the indirect-injection chain, and redaction before\n  embedding. Do not use it for choosing an embedding model, for retrieval quality, or for the shape of the\n  corpus itself (padosoft-rag-knowledge-base).\nlicense: MIT\ncompatibility: >-\n  Stack-agnostic. Written for a multi-tenant assistant with pluggable connectors and a vector store; the\n  reasoning applies to any pipeline that turns somebody else's documents into grounding.\nmetadata:\n  version: 0.1.0\n  author: Padosoft\n  summary: The ingestion contract records what a document is, not where its authority comes from.\n  profiles: ai, api\n  scope: project\n  repository: https:\u002F\u002Fgithub.com\u002Fpadosoft\u002Fskills\n  keywords: rag, ingestion, connectors, acl, oversharing, provenance, prompt injection, pii, redaction\n---\n\n# RAG ingestion security\n\nIngestion looks like a data problem and is an authorisation problem. **The contract records what a document\n*is* — path, title, type, tenant — and not where its authority comes from.** There is nowhere to put the\nfact that it was shared with three people, so it arrives readable by everyone with access to the corpus.\n\nThat is oversharing, and it is the headline failure of enterprise assistant rollouts. The second failure\nshares its cause: nothing records **who wrote** the content either.\n\n---\n\n## 1. Mirror the source's permissions, or say that you did not\n\nA connector reports **what the source said**; it never decides who may read. The distinction is the design.\n\n```text\nprincipals (type + external id + allow\u002Fdeny)\ninherits_from_parent: bool      folder or space inheritance\ncomplete: bool                  false = the source truncated the list\n```\n\nThe host maps external principals to internal subjects — by verified address, directory link, group\nmapping. That resolution depends on directory state the connector does not have, which is why it is not the\nconnector's job.\n\n**And it will sometimes fail**: an external collaborator, a group the directory does not know, a truncated\nlist.\n\n## 2. Unmapped access fails closed. This is the load-bearing rule\n\n**A document whose access could not be fully mapped must not fall back to corpus-wide visibility.** That\nfallback is exactly the bug this whole design exists to remove, and it is the tempting shortcut every time.\n\nStore it in a restricted state — readable only by the owner of the connector installation — until an\noperator triages it. **This makes some documents invisible that were visible yesterday. That is the point,\nnot a regression**, and saying so up front is what keeps it from being reverted in the first week.\n\n## 3. Mirrored permissions have to be revoked, not only granted\n\n- Keep the **origin** of every permission row: mirrored from the source, or set by hand. A re-sync\n  reconciles the mirrored ones and must not destroy an operator's manual grant.\n- **When a source removes a share, the reconciliation deletes the mirrored row.** A mirror that only ever\n  adds is a slow leak, and it leaks in the direction nobody is watching.\n- Record an unresolved principal **as an unresolved principal** rather than discarding it. The row is the\n  evidence that access was intended for somebody, and it is what an operator triages later.\n- Deny wins over allow, wherever both apply.\n\n## 4. Authorisation belongs inside the query\n\nThe retrieval path is a hot path, and it does not call your policy object. Push the allow-list into the\nquery — the same predicate the rest of the application uses — so retrieval gets authorisation **for free\nand by construction** rather than as a filter somebody remembered to add after the results came back.\n\nA decision made after retrieval is a decision that can be skipped by the next code path that queries\ndirectly. See **`padosoft-tenant-isolation`**.\n\n## 5. Provenance: who wrote this, and do we trust them?\n\n```text\ntrusted-internal · untrusted-external · machine-generated\n```\n\nMost connectors bring in content written inside the organisation. **A mailbox connector brings in content\nwritten by anyone who can send an email.** That message becomes a document, the document becomes chunks, the\nchunks become grounding — and the same platform exposes tools an agent can call. That is a complete\n**indirect-injection chain**: attacker-controlled text reaching a tool-calling context, with no boundary\nbetween the two ends.\n\n- **Ship the label before the enforcement.** A provenance column and a contract method are cheap, and they\n  immediately answer a question nobody can answer today: *how much of our corpus is externally authored?*\n  Labels without enforcement are still worth having — they make the enforcement testable before it exists.\n- **Then enforce the asymmetry**: an untrusted chunk may be **quoted in an answer** and must **never\n  influence a tool call**. A tool invocation whose arguments derive from untrusted grounding needs an\n  explicit policy or a human confirmation. See **`padosoft-agent-host-boundaries`**.\n- **Provenance is not curation tier, and collapsing them loses information.** Curation asks *has a human\n  vouched for this?*; provenance asks *who wrote it?* A human-approved page summarising an external email is\n  **both** approved and externally authored, and both facts matter at different moments.\n\n## 6. Personal data: redact before embedding\n\nOnce a value is in the vector store it is in every backup, every index and every answer. So the redaction\nruns **at ingestion, before the embedding**, not at display time.\n\n- **Deterministic surrogates**, so the same value always yields the same token and search still works — a\n  search for a person's ticket still matches.\n- **Two strategies**: one-way masking, and reversible tokenisation with a vault. The vault is **per tenant,\n  with a per-tenant salt**, so the same value in two tenants produces different tokens and nothing\n  correlates across them.\n- **One redaction core for every ingestion path.** A pipeline path and a direct path that redact separately\n  will diverge, and the contract *no raw personal data in the index* has to be path-independent.\n- **Redact the chunk, not the source.** The original stays as the idempotency anchor and keeps its metadata\n  parseable — see **`padosoft-rag-knowledge-base`**.\n- **A dry run uses the one-way strategy**, so a preview never writes to the vault.\n- **Re-identification is a privileged, audited operation**, and stricter still when exposed as a tool a\n  model can call. Every unmask and every denied attempt writes an immutable audit row that records **counts,\n  never the values**.\n\n## 7. Extending the contract without stranding anybody\n\nThe ingestion contract is public API for every connector, including the ones other people wrote.\n\n- **Add optional capability interfaces**, never a required argument. A connector implementing neither\n  behaves exactly as before.\n- Carry the new data in the **extension channel that already exists** — the metadata map — written and read\n  through helpers, so no call site handles the key by hand.\n- **An optional trailing parameter is not additive for implementers.** Callers are fine; a class that\n  *implements* the interface with the old signature fails at declaration time, before any of its code runs.\n  That is the compatibility you were trying to preserve, broken by the mechanism chosen to preserve it. See\n  **`padosoft-contract-changes`**.\n\n---\n\n## Gotchas\n\n- **Nothing in the code is wrong when oversharing happens.** The contract simply had nowhere to put the\n  permission, so no reviewer sees a mistake.\n- **A connector that reads permissions but ingests them as \"allow everyone\" on failure is worse than one\n  that reads none**, because it looks solved.\n- **The retrieval path is the one that skips the policy**, every time, in every codebase.\n- **A summary of untrusted content inherits its provenance**, and a human approving the summary does not\n  change who wrote the original.\n- **Determinism in redaction is a privacy trade**: identical surrogates make search work and make frequency\n  analysis possible. Per-tenant salting is what keeps it acceptable.\n- **The audit trail of a re-identification must not contain what was re-identified.**\n\n## Checklist\n\n- [ ] The connector reports principals and completeness; it makes no access decision\n- [ ] Unmapped or truncated access fails **closed**, into a restricted state with an owner and a triage path\n- [ ] Permission rows carry their origin; re-sync reconciles mirrored rows without destroying manual ones\n- [ ] Revocation implemented and tested, not only granting\n- [ ] Authorisation enforced inside the retrieval query\n- [ ] Provenance tier assigned at ingestion; labels shipped before enforcement\n- [ ] Untrusted grounding may be quoted, never influence a tool call without policy or confirmation\n- [ ] Provenance kept separate from curation tier\n- [ ] Redaction before embedding, deterministic, one core for every path, chunk not source\n- [ ] Vault per tenant with a per-tenant salt; re-identification privileged and audited without values\n- [ ] Contract extended by optional capability interfaces; no implementer stranded\n\n## Final report\n\n```\nConnectors: \u003Cn> — with ACL mirroring: \u003Cn> · with provenance: \u003Cn>\nUnmapped policy: \u003Cwhat happens> · documents currently restricted-unmapped: \u003Cn>\nRevocation: implemented | not (leak direction: \u003C…>)\nEnforcement: in-query | post-filter (where)\nProvenance: labels \u003Cshipped?> · enforcement \u003Cphase> · untrusted corpus share: \u003C%>\nRedaction: strategy \u003C…> · vault \u003Cper-tenant salt?> · paths covered: \u003Clist>\nContract changes: \u003Cwhat was added, and how implementers stay valid>\n```\n","# RAG ingestion security\n\nIngestion looks like a data problem and is an authorisation problem. **The contract records what a document\n*is* — path, title, type, tenant — and not where its authority comes from.** There is nowhere to put the\nfact that it was shared with three people, so it arrives readable by everyone with access to the corpus.\n\nThat is oversharing, and it is the headline failure of enterprise assistant rollouts. The second failure\nshares its cause: nothing records **who wrote** the content either.\n\n---\n\n## 1. Mirror the source's permissions, or say that you did not\n\nA connector reports **what the source said**; it never decides who may read. The distinction is the design.\n\n```text\nprincipals (type + external id + allow\u002Fdeny)\ninherits_from_parent: bool      folder or space inheritance\ncomplete: bool                  false = the source truncated the list\n```\n\nThe host maps external principals to internal subjects — by verified address, directory link, group\nmapping. That resolution depends on directory state the connector does not have, which is why it is not the\nconnector's job.\n\n**And it will sometimes fail**: an external collaborator, a group the directory does not know, a truncated\nlist.\n\n## 2. Unmapped access fails closed. This is the load-bearing rule\n\n**A document whose access could not be fully mapped must not fall back to corpus-wide visibility.** That\nfallback is exactly the bug this whole design exists to remove, and it is the tempting shortcut every time.\n\nStore it in a restricted state — readable only by the owner of the connector installation — until an\noperator triages it. **This makes some documents invisible that were visible yesterday. That is the point,\nnot a regression**, and saying so up front is what keeps it from being reverted in the first week.\n\n## 3. Mirrored permissions have to be revoked, not only granted\n\n- Keep the **origin** of every permission row: mirrored from the source, or set by hand. A re-sync\n  reconciles the mirrored ones and must not destroy an operator's manual grant.\n- **When a source removes a share, the reconciliation deletes the mirrored row.** A mirror that only ever\n  adds is a slow leak, and it leaks in the direction nobody is watching.\n- Record an unresolved principal **as an unresolved principal** rather than discarding it. The row is the\n  evidence that access was intended for somebody, and it is what an operator triages later.\n- Deny wins over allow, wherever both apply.\n\n## 4. Authorisation belongs inside the query\n\nThe retrieval path is a hot path, and it does not call your policy object. Push the allow-list into the\nquery — the same predicate the rest of the application uses — so retrieval gets authorisation **for free\nand by construction** rather than as a filter somebody remembered to add after the results came back.\n\nA decision made after retrieval is a decision that can be skipped by the next code path that queries\ndirectly. See **`padosoft-tenant-isolation`**.\n\n## 5. Provenance: who wrote this, and do we trust them?\n\n```text\ntrusted-internal · untrusted-external · machine-generated\n```\n\nMost connectors bring in content written inside the organisation. **A mailbox connector brings in content\nwritten by anyone who can send an email.** That message becomes a document, the document becomes chunks, the\nchunks become grounding — and the same platform exposes tools an agent can call. That is a complete\n**indirect-injection chain**: attacker-controlled text reaching a tool-calling context, with no boundary\nbetween the two ends.\n\n- **Ship the label before the enforcement.** A provenance column and a contract method are cheap, and they\n  immediately answer a question nobody can answer today: *how much of our corpus is externally authored?*\n  Labels without enforcement are still worth having — they make the enforcement testable before it exists.\n- **Then enforce the asymmetry**: an untrusted chunk may be **quoted in an answer** and must **never\n  influence a tool call**. A tool invocation whose arguments derive from untrusted grounding needs an\n  explicit policy or a human confirmation. See **`padosoft-agent-host-boundaries`**.\n- **Provenance is not curation tier, and collapsing them loses information.** Curation asks *has a human\n  vouched for this?*; provenance asks *who wrote it?* A human-approved page summarising an external email is\n  **both** approved and externally authored, and both facts matter at different moments.\n\n## 6. Personal data: redact before embedding\n\nOnce a value is in the vector store it is in every backup, every index and every answer. So the redaction\nruns **at ingestion, before the embedding**, not at display time.\n\n- **Deterministic surrogates**, so the same value always yields the same token and search still works — a\n  search for a person's ticket still matches.\n- **Two strategies**: one-way masking, and reversible tokenisation with a vault. The vault is **per tenant,\n  with a per-tenant salt**, so the same value in two tenants produces different tokens and nothing\n  correlates across them.\n- **One redaction core for every ingestion path.** A pipeline path and a direct path that redact separately\n  will diverge, and the contract *no raw personal data in the index* has to be path-independent.\n- **Redact the chunk, not the source.** The original stays as the idempotency anchor and keeps its metadata\n  parseable — see **`padosoft-rag-knowledge-base`**.\n- **A dry run uses the one-way strategy**, so a preview never writes to the vault.\n- **Re-identification is a privileged, audited operation**, and stricter still when exposed as a tool a\n  model can call. Every unmask and every denied attempt writes an immutable audit row that records **counts,\n  never the values**.\n\n## 7. Extending the contract without stranding anybody\n\nThe ingestion contract is public API for every connector, including the ones other people wrote.\n\n- **Add optional capability interfaces**, never a required argument. A connector implementing neither\n  behaves exactly as before.\n- Carry the new data in the **extension channel that already exists** — the metadata map — written and read\n  through helpers, so no call site handles the key by hand.\n- **An optional trailing parameter is not additive for implementers.** Callers are fine; a class that\n  *implements* the interface with the old signature fails at declaration time, before any of its code runs.\n  That is the compatibility you were trying to preserve, broken by the mechanism chosen to preserve it. See\n  **`padosoft-contract-changes`**.\n\n---\n\n## Gotchas\n\n- **Nothing in the code is wrong when oversharing happens.** The contract simply had nowhere to put the\n  permission, so no reviewer sees a mistake.\n- **A connector that reads permissions but ingests them as \"allow everyone\" on failure is worse than one\n  that reads none**, because it looks solved.\n- **The retrieval path is the one that skips the policy**, every time, in every codebase.\n- **A summary of untrusted content inherits its provenance**, and a human approving the summary does not\n  change who wrote the original.\n- **Determinism in redaction is a privacy trade**: identical surrogates make search work and make frequency\n  analysis possible. Per-tenant salting is what keeps it acceptable.\n- **The audit trail of a re-identification must not contain what was re-identified.**\n\n## Checklist\n\n- [ ] The connector reports principals and completeness; it makes no access decision\n- [ ] Unmapped or truncated access fails **closed**, into a restricted state with an owner and a triage path\n- [ ] Permission rows carry their origin; re-sync reconciles mirrored rows without destroying manual ones\n- [ ] Revocation implemented and tested, not only granting\n- [ ] Authorisation enforced inside the retrieval query\n- [ ] Provenance tier assigned at ingestion; labels shipped before enforcement\n- [ ] Untrusted grounding may be quoted, never influence a tool call without policy or confirmation\n- [ ] Provenance kept separate from curation tier\n- [ ] Redaction before embedding, deterministic, one core for every path, chunk not source\n- [ ] Vault per tenant with a per-tenant salt; re-identification privileged and audited without values\n- [ ] Contract extended by optional capability interfaces; no implementer stranded\n\n## Final report\n\n```\nConnectors: \u003Cn> — with ACL mirroring: \u003Cn> · with provenance: \u003Cn>\nUnmapped policy: \u003Cwhat happens> · documents currently restricted-unmapped: \u003Cn>\nRevocation: implemented | not (leak direction: \u003C…>)\nEnforcement: in-query | post-filter (where)\nProvenance: labels \u003Cshipped?> · enforcement \u003Cphase> · untrusted corpus share: \u003C%>\nRedaction: strategy \u003C…> · vault \u003Cper-tenant salt?> · paths covered: \u003Clist>\nContract changes: \u003Cwhat was added, and how implementers stay valid>\n```\n",{"name":5,"license":14,"metadata":23,"description":7,"compatibility":31},{"scope":24,"author":25,"summary":26,"version":27,"keywords":28,"profiles":29,"repository":30},"project","Padosoft","The ingestion contract records what a document is, not where its authority comes from.","0.1.0","rag, ingestion, connectors, acl, oversharing, provenance, prompt injection, pii, redaction","ai, api","https:\u002F\u002Fgithub.com\u002Fpadosoft\u002Fskills","Stack-agnostic. Written for a multi-tenant assistant with pluggable connectors and a vector store; the reasoning applies to any pipeline that turns somebody else's documents into grounding.","main","8ff266762b08cab7f5be5551fa221309dc9db48b","skills\u002Fpadosoft-rag-ingestion-security",0,"https:\u002F\u002Fgithub.com\u002Fpadosoft","https:\u002F\u002Fwww.padosoft.com","Software House","2026-10-19 01:02:17",{"source":8,"skillName":5,"commands":41,"prompt":51,"zipAvailable":52},[42,45,48],{"tool":43,"command":44},"npx","npx skills add padosoft\u002Fskills --skill padosoft-rag-ingestion-security",{"tool":46,"command":47},"bunx","bunx skills add padosoft\u002Fskills --skill padosoft-rag-ingestion-security",{"tool":49,"command":50},"pnpm","pnpm dlx skills add padosoft\u002Fskills --skill padosoft-rag-ingestion-security","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fpadosoft\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002Fpadosoft-rag-ingestion-security\u002FSKILL.md",true,[54],{"path":55,"resourceType":56,"resourceTypeDesc":57,"sizeBytes":58,"sourceHtmlUrl":59},"evals\u002Fqueries.json","REFERENCE","参考文档",1560,"https:\u002F\u002Fgithub.com\u002Fpadosoft\u002Fskills\u002Fblob\u002Fmain\u002Fskills\u002Fpadosoft-rag-ingestion-security\u002Fevals\u002Fqueries.json",[61,68,74,81,88,94],{"skillNo":62,"name":63,"nameZh":6,"description":64,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":9,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":65,"categories":66,"keywords":67,"publishedAt":18},"01M7CZNKCTMWK7F2J6KKPAR8YM","report-opus","Writes the weekly status report from the team's tickets. Use when the user asks for the weekly report or a status summary. Do not use it for incident reports.","https:\u002F\u002Fgithub.com\u002Fpadosoft\u002Fskills\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fskill-audit\u002Fproject\u002Fdot-claude\u002Fskills\u002Freport-opus\u002FSKILL.md",[],[],{"skillNo":69,"name":70,"nameZh":6,"description":64,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":9,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":71,"categories":72,"keywords":73,"publishedAt":18},"01M7CZNFEN8XK6S8NC6KGREYGS","report-haiku","https:\u002F\u002Fgithub.com\u002Fpadosoft\u002Fskills\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fskill-audit\u002Fproject\u002Fdot-claude\u002Fskills\u002Freport-haiku\u002FSKILL.md",[],[],{"skillNo":75,"name":76,"nameZh":6,"description":77,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":9,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":78,"categories":79,"keywords":80,"publishedAt":18},"01M7CZND182TKGGM7R4134ND00","kitt-stub","Builds the guided-tour reference for an admin page. Use when the user asks for the tour reference. Trigger \"tour\", \"reference\", \"guided tour\".","https:\u002F\u002Fgithub.com\u002Fpadosoft\u002Fskills\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fskill-audit\u002Fproject\u002Fdot-claude\u002Fskills\u002Fkitt-stub\u002FSKILL.md",[],[],{"skillNo":82,"name":83,"nameZh":6,"description":84,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":9,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":85,"categories":86,"keywords":87,"publishedAt":18},"01M7CZN9F2A6BFTFW5RSRNAGB1","good-skill","Generates the release notes of a project from its merged pull requests. Use this skill when the user asks for release notes, a changelog entry or a summary of what shipped since the last tag. Do not use it to write commit messages, that is the job of a commit skill.","https:\u002F\u002Fgithub.com\u002Fpadosoft\u002Fskills\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fskill-audit\u002Fproject\u002Fdot-claude\u002Fskills\u002Fgood-skill\u002FSKILL.md",[],[],{"skillNo":89,"name":90,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":9,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":91,"categories":92,"keywords":93,"publishedAt":18},"01M7CZN6BJASPD4DH8DFR63GQF","broken-frontmatter","https:\u002F\u002Fgithub.com\u002Fpadosoft\u002Fskills\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fskill-audit\u002Fproject\u002Fdot-claude\u002Fskills\u002Fbroken-frontmatter\u002FSKILL.md",[],[],{"skillNo":95,"name":96,"nameZh":6,"description":97,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":9,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":98,"categories":99,"keywords":100,"publishedAt":18},"01M7CZN498TRNC30GSYMMAZNXM","Bad_Skill","I can help you with reports.","https:\u002F\u002Fgithub.com\u002Fpadosoft\u002Fskills\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fskill-audit\u002Fproject\u002Fdot-claude\u002Fskills\u002Fbad-skill\u002FSKILL.md",[],[]]