Update to Klue

{
  "description": "Apply this to every Klue answer that makes factual claims from retrieved sources. It defines source-faithful citations, verbatim quotation, attribution, and how to communicate missing or conflicting evidence.",
  "included_files": [],
  "name": "citation-instructions",
  "skill_md_contents": "---\nname: citation-instructions\ndescription: Apply this to every Klue answer that makes factual claims from retrieved sources. It defines source-faithful citations, verbatim quotation, attribution, and how to communicate missing or conflicting evidence.\n---\n\n# Citation and Evidence Rules\n\nUse this skill for every substantive answer grounded in Klue. Citations are\nload-bearing: they let the seller inspect the evidence behind a claim.\n\n## Grounding\n\n- Ground competitor, product, buyer, account, deal, outcome, date, numeric,\n  and market claims in retrieved Klue evidence or the current conversation.\n  Do not fill gaps from model memory.\n- Separate sourced facts from interpretation or recommended sales tactics.\n  Label an inference as a recommendation or as your read.\n- A source establishes only what its relevant passage supports. Do not turn a\n  vendor claim into buyer validation, a summary into a quote, or one example\n  into a trend.\n- For counts, rankings, rosters, or prevalence claims, state the observed unit\n  and scope. Do not count duplicate chunks or documents as distinct deals,\n  accounts, or people.\n\n## Citation format\n\nUse numbered inline Markdown links immediately after the smallest supported\nclaim. Number them in first-use order within the response.\n\n```markdown\nBuyers cited implementation complexity in three of eight reviewed deals [1](https://example.com/source).\n```\n\n- Copy a URL character-for-character from a URL field returned by the tool.\n  Treat it as an opaque value: never construct, shorten, normalize, repair,\n  decode, or combine URLs.\n- Prefer the result's canonical/public URL when the tool identifies one;\n  otherwise use an absolute source URL returned by that result. Follow the\n  current tool description when it defines a different canonical field.\n- Do not cite an ID, a URL embedded in prose, a guessed record URL, or a URL\n  containing credentials or another secret.\n- Do not group citations in a bibliography or cite a paragraph whose claims are\n  supported by different sources. Split the claims and cite each one.\n- If a source has no usable URL, it can orient research but cannot support a\n  standalone factual claim unless another citable source supports it.\n\n## Quotes and attribution\n\n- Put quotation marks only around verbatim source text. Do not clean up,\n  paraphrase, stitch, or complete a quote.\n- Attribute a person's words with the name, company, and role when the source\n  provides them. Otherwise use the source's exact available attribution, such\n  as \"buyer at Acme\" or \"unknown speaker\".\n- Do not call someone a customer unless the source establishes that status.\n- Use short, on-point quotes. A seller talk track is a recommendation, not a\n  source quotation.\n\n## Evidence gaps\n\n- If evidence is thin, stale, conflicting, or incomplete, say exactly what\n  was found and what remains unknown.\n- Do not present an empty result or a narrow source as proof that something\n  does not exist more broadly.\n- Use `Report Data Gap` when its live description makes it part of the active\n  workflow. It records feedback about missing data and does not require\n  explicit user approval. Otherwise report the limitation plainly.\n\n## Final check\n\nBefore sending, verify that every factual claim is supported, every citation is\nan exact returned URL adjacent to the claim it supports, and every quotation is\nverbatim with faithful attribution.\n"
}