evidence-citations
Anchor every claim in an explanation to file:line evidence you actually opened, and mark the difference between what you verified and what you inferred.
- License 许可
- Apache-2.0
- 最近更新
- 2026-11-15
- Stars
- ★ 827
- Forks
- 573
01 前言 Front Matter
- name
- evidence-citations
- description
- Anchor every claim in an explanation to file:line evidence you actually opened, and mark the difference between what you verified and what you inferred.
03 SKILL.md 正文
Evidence and citations
Every non-obvious claim carries a path/to/file.ext:LINE anchor. This is also
what closes a path in an understanding run — a finding recorded without one is
downgraded to an open conjecture, so an uncited explanation is not merely weak,
it does not count.
- Cite the line that actually shows the behaviour: the write, the query, the branch. Citing the type definition proves the entity exists, which was never in question.
- Never cite a file you did not open. A plausible path is the easiest thing in this job to invent and the most expensive thing for a reader to discover is wrong.
- Prefer two anchors that disagree over one that is tidy. "Written here, but also mutated here" is the finding.
- When you searched and found nothing — no writer, no reader, no caller — say so explicitly and give the search you ran. An unreferenced table is a real and useful result, not a failure to look.
Separate verified from inferred in the document itself. Mark inference as inference ("appears to be", "no caller found in this repo") and keep it out of the summary lines a reader will quote. In the HTML, render anchors in monospace so they are visibly checkable, and keep the repo, branch and commit in the header so a stale citation can be dated rather than merely doubted.
04 同仓库更多 Skills
whoami
Show current Xyne Claw auth identity.
sessions
List Xyne Claw sessions.
run
Run a Xyne Claw agent.
login
Log the user into Xyne Claw.
agents
List available Xyne Claw agents.
explanation-document
Structure a codebase/schema explanation as a self-contained HTML document ordered for a READER — mental model and flows first, deep per-component reference last — not in the order you explored it.