用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt;运行前请先检查来源。
npx skills add https://github.com/jmagly/aiwg --skill verify-citations命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地?可下载 SkillsMP 当前能够提供的文件。
正在显示 SKILL.md
基于 SOC 职业分类
| namespace | aiwg |
| name | verify-citations |
| platforms | ["all"] |
| description | Verify all citations in a document against the research corpus |
| commandHint | {"category":"research-quality"} |
Validate that all citations, references, and factual claims are backed by actual sources in the research corpus.
When invoked, perform systematic citation verification:
Parse Target File
@.aiwg/research/sources/* references@.aiwg/research/findings/* referencesREF-XXX inline references10.XXXX/...)Author et al., YYYY)[<artifact_id>@<content_hash>, L<n>-<n>] and path-only fallback [<artifact_id>, L<n>-<n>]Verify Each Citation
content_hash present, recompute current hash and compare (see RLM Citation Verification below)Detect Hallucinations
Check GRADE Compliance
Generate Report
Auto-Fix Mode (--fix)
[STALE: hash mismatch]; do not auto-rewrite (the agent that emitted them must decide whether the new content still supports the claim)When the target file contains RLM-format citations (#1205, #1223), verify-citations validates them against the current state of the cited artifacts.
The skill recognizes two inline forms (matching the schema at agentic/code/addons/rlm/schemas/citation-tuple.json):
[<artifact_id>@<content_hash>, <lines>] — full provenance with content hash
[src/auth/login.ts@a8f9c2d4e5f60718, L42-58][<artifact_id>, <lines>] — citation without content hash (artifact wasn't in the index when emitted)
[src/legacy/auth.py, L88]The lines field is optional in both forms.
For versioned RLM citations:
<artifact_id> to a project-relative file pathsrc/artifacts/index-builder.ts:243)content_hash:
--strict)For path-only fallback RLM citations:
<artifact_id> to a project-relative file path--strict)Stale verdict means the file content has changed since the citation was emitted — the claim may no longer be supported by the cited region. The simple hash-mismatch heuristic doesn't distinguish whitespace edits from semantic changes; richer semantic-drift detection is the responsibility of aiwg index doctor --rlm-audit (#1208).
The hash routine MUST match the index implementation exactly so citations emitted by rlm-batch round-trip cleanly:
sha256(file_contents).hexdigest()[0:16]
See src/artifacts/index-builder.ts:243 for the canonical implementation. Any verification logic that diverges from this convention will produce false stale verdicts.
When RLM citations are present, the report adds a new section:
RLM Citations: 12 total
✓ Fresh: 9
⚠ Stale: 2
⚠ Un-versioned: 1
✗ Missing: 0
STALE citations:
- [src/auth/login.ts@a8f9c2d4e5f60718, L42-58]
Current hash: f3b2c1d4e5a60718
File modified since citation emitted
UN-VERSIONED citations:
- [src/legacy/auth.py, L88]
File exists but no content hash; consider running through aiwg index
--rlm: opt-in to RLM-only mode — skip @-mention research-corpus checks. Useful when verifying RLM-aggregated outputs in CI without research-corpus context.--strict: existing flag — STALE and UN-VERSIONED RLM citations escalate from warnings to errors and produce non-zero exit.[file-path] - File to verify (default: all .md files in current directory)--strict - Treat warnings as errors (including STALE and UN-VERSIONED RLM citations)--fix - Automatically fix GRADE violations and remove hallucinated citations (does not rewrite RLM citations — see #6 above)--report - Save report to .aiwg/reports/citation-verification.md--corpus-only - Only check @-mention references, skip author-year patterns--rlm (#1225) - Verify RLM citations only; skip research-corpus checks