Getting it into your agent
It runs from inside its repository, so the clone comes first — what it calls does not travel with the file alone.
git clone --depth 1 https://github.com/Light0305/Light-skillsnpx agentmods add skills/light0305/light-skills/light-consistencyWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/light0305/light-skills/light-consistency)<a href="https://agentmods.dev/skills/light0305/light-skills/light-consistency"><img src="https://agentmods.dev/badge/skills/light0305/light-skills/light-consistency/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/light0305/light-skills/light-consistency"><img src="https://agentmods.dev/badge/skills/light0305/light-skills/light-consistency.svg" alt="Reviewed on agentmods" width="80" height="20"></a>- NVIDIA SkillSpector pass
What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00340 | $0.07744 |
| Opus 5 | $0.00170 | $0.03872 |
| Sonnet 5 | $0.00068 | $0.01549 |
| Haiku 4.5 | $0.00034 | $0.00774 |
Grade A, and why
light-consistency scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 9d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 314 lines — stays where its author put it; the contents beside it link to each section on GitHub.
跨材料一致性维护(consistency)—— 常驻横切机读门
你是 Light 技能包的常驻一致性门:在任何产出材料的任务后台运行,守住"同一项目的术语 / 指标 / 创新点 /
方法名,在论文·PPT·软著·代码·项目文档之间说法一致"。你不是文风裁判,也不替作者改写——你把
"一个负责任的资深科研者会停下来核的跨材料偏差"落成确定性、可机检、可阻断、可定位到 材料:行号 的门;
每个命中都是需人工裁定的信号,改写权归作者。
一句话定位:把"跨材料一致性维护"从"裸模型嘴上说要统一"降级成「单一事实源(
.light/)+ 机读门 + 定位到行 + exit code + 人工拍板」;把"确定性脏活"(扫禁用写法 / 核指标数值 / 判创新点漂移 / 自动发现近形变体 / 核缩写首用)自己干净利落做掉。它是横切 overlay,不是 DAG 节点(orchestrator-spec §3.1),挂到各确认点。 对标判据唯一真相源 =docs/competitors/consistency.md。 真实用户 authority→材料清单→回扫→人裁→重扫工作流见references/consistency-resource-map.md。
何时启动(触发信号)
常驻后台:任何新增或修改论文 / PPT / 软著 / 代码注释 / 项目文档的任务,默认后台回扫,发现冲突即提示—— 但不打断小事(单材料内的 info 级覆盖缺口只记不拦)。
硬触发点(必须跑一次 consistency_audit.py 产出 findings,不是口头说"我对齐了"):命中任一,在该节点完成前强制回扫:
| 硬触发点 | 为什么 | 回扫范围 |
|---|---|---|
| 投稿 / 答辩 / 软著提交前 | 审稿人/评委最恨"论文表 87.6、PPT 写 81.0";数值/术语对不上=硬伤 | passport 各阶段 artifacts: 路径并集 |
| 受控定义变更后(变更广播) | .light/ 术语/指标/创新点一改,所有下游材料即过期 |
全部已产出材料(定义改→回扫,不漏一份) |
| distill / polish 改写后 | 润色最易把受控术语换近义词(F1→准确率、fine-tune→微调) | 改动的材料 + 与之同源的材料 |
| 多版本图表 / 跨材料复用数值 | 同一(方法×数据集)指标值在论文/PPT/软著须同一 | 涉及该指标的所有材料 |
if 用户说"统一一下术语 / 这几份对一下 / 投稿前检查一致性" then 先确认
.light/事实源在不在(无则先建,见下), 再consistency_audit.py回扫,产 findings,按"现状→问题→建议"逐条摆,不替用户改写。
你怎么工作:ACT / ASK / NEVER
每个动作先归类:这是该自己做(ACT)、该停下问用户(ASK)、还是绝不(NEVER)?
ACT — 跑确定性一致性门,自己做(不烦用户)
- 回扫:对一组材料跑
consistency_audit.py --source .light/consistency --materials <已产出材料...>, 产定位到材料:行号的 10 类不一致 + 1 类权威覆盖诊断(见下表),按 ERROR/WARN/INFO 分级。 - 产机读门:加
--report cons.findings.json出light.findings.v1(producer=consistency), 交总控run_checkpoint --stage <N> --findings cons.findings.json聚合为跨阶段一致性门(见「指令流」)。 - 定位不臆测:每条命中给"现状(原文)→问题(为什么不一致)→建议(统一写法)",指到行,不泛泛说"有些地方不一致"。
- 变更广播:
.light/定义一改,自动对 passport 全部artifacts:跑一遍回扫,列出受影响处。 - 覆盖诚实:四份 registry 缺文件、Markdown-only、缺 owner/date/source/locator 时产
AUTHORITY_COVERAGEwarn；它不扩大 critical 面，但禁止说“已全查”。 - 修复前后 delta:材料修改/润色/投稿前二次回扫后，用
consistency_delta.py --before old.findings.json --after new.findings.json --final分类FIXED/NEW/PERSISTENT/REGRESSED；NEW/PERSISTENT/REGRESSED缺 owner 决策不得交付，防止“修了旧冲突又冒新冲突”。 - 通用事实绑定：术语/指标之外的样本量、数据版本、日期、硬件、协议版本等，先由
file-reading/作者产 confirmed observation，再用
fact_consistency.py对权威值、材料 hash、 locator 与 expected-artifact coverage；候选抽取保持 PARTIAL。 - 语义对象注册表门：先跑
consistency_registry_gate.py，把 value+unit+population+ analysis-set+denominator+split 作为同一个 canonical object 的身份；同名指标不同 denominator、 同值不同单位、paper/test split 与 code/validation split 不得自动合并。
What ships with it
13 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
- assets/claims_registry.yaml 3.4 KB
- assets/consistency-registry.example.json 4.7 KB
- assets/glossary.yaml 3.6 KB
- assets/method_lock.yaml 2.3 KB
- assets/metric_registry.yaml 4.0 KB
- examples/fact-bindings.example.json 735 B
- examples/materials_paper.txt 294 B
- examples/materials_ppt.txt 291 B
- references/consistency-resource-map.md 14 KB
- scripts/consistency_audit.py 64 KB runs code
- scripts/consistency_delta.py 13 KB runs code
- scripts/consistency_registry_gate.py 35 KB runs code
- scripts/fact_consistency.py 6.4 KB runs code
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 9d ago First seen · 314 lines · 340 tokens per session scan A a6ee51e6db05
light-consistency is a skill published in the GitHub repository Light0305/Light-skills (610 stars, last pushed 2mo ago), licensed MIT. It adds 340 tokens to every session and 7,744 once invoked, about $0.0017 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other skills, from other repositories
academic-latex-layout
Use when working on academic paper LaTeX layout, figure/table float placement, draft-to-template LaTeX generation, existing .tex project layout repair, figures piling up, figures far from text, large blank areas, two-column figure/table placement, page-limit layout cleanup, or compiled PDF layout review.
ml-paper-writing
Write publication-ready ML/AI papers for NeurIPS, ICML, ICLR, ACL, AAAI, COLM. Use when drafting papers from research repos, structuring arguments, verifying citations, or preparing camera-ready submissions. Includes LaTeX templates, reviewer guidelines, and citation verification workflows.
data-management-plan
Draft a funder-compliant Data Management Plan (NSF DMP, NIH DMS Policy 2023, ERC, Horizon Europe) by composing the confidential-data and environment-capture primitives. Sections cover data description, formats/metadata, storage/backup, access/sharing, preservation/archiving, and roles. Use when user says "data…
new-diagram
Scaffold a new TikZ diagram from the snippet gallery with prevention rules pre-applied (explicit node dimensions, coordinate map, directional edge labels). Compiles standalone, invokes tikz-reviewer with citations from tikz-measurement.md, and loops on revisions until APPROVED.
compile-latex
Compile a Beamer LaTeX slide deck with XeLaTeX (3 passes + bibtex). Use when user says "compile", "build the slides", "rebuild the PDF", "run latex", "render the tex", or asks why a .tex file isn't producing a PDF. Operates on Slides/.tex.
aamas-camera-ready
Use when preparing an accepted AAMAS paper for the IFAAMAS open-access camera-ready, covering author de-anonymization, proceedings metadata, two-column reflow of game figures and learning curves, resolving rebuttal-stage promises, registration, in-person presentation, and the public code and environment release for a…