Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add AbyssCN/oh-my-dag --skill grillgit clone --depth 1 https://github.com/AbyssCN/oh-my-dagWrote 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/abysscn/oh-my-dag/grill)<a href="https://agentmods.dev/skills/abysscn/oh-my-dag/grill"><img src="https://agentmods.dev/badge/skills/abysscn/oh-my-dag/grill/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/abysscn/oh-my-dag/grill"><img src="https://agentmods.dev/badge/skills/abysscn/oh-my-dag/grill.svg" alt="Reviewed on agentmods" width="80" height="20"></a>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 | $0.00116 | $0.02130 |
| Opus 5.5 | $0.00046 | $0.00852 |
| Sonnet 5.5 | $0.00023 | $0.00426 |
| Haiku 4.5 | $0.00012 | $0.00213 |
Grade A, and why
grill 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 5d 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 — 79 lines — stays where its author put it; the contents beside it link to each section on GitHub.
/grill — 锁契约前的对抗式审问
对应 pi TUI 的 /grill(plan mode 审议纪律的独立版),吸收 Aalto grill-me 的逼问纪律。审议期间只讨论不改代码(无代码闸,靠自律)。产物不是「共识散文」,是决策记录表 + 就地落盘的裁决,直接喂 /contract。
When to use
一个 plan / 设计方向还停在「大概这样」,锁进 /contract 或交执行之前:
- 新模块选型未定;多个方案僵持,想被逼着把每条分支走到底;自觉计划有没想清的洞,要个蓝军。
不用于:已定方案的实装(直接交 Opus 子 agent 或 omd-lite run)/ 代码审查(/omd-review)/ 根因调试。
五条铁律(承 grill-me)
姿态 = co-operator 不是应答机:抗中庸(从 first-principles 切入,不给平均方案)+ 永不主动停(每轮收尾带前进动作,不停在「你看怎么样?」)。其余「主动给见解 / 沿决策树 / 能自查先自查」已在下面铁律里,不重列。
- 问 owner 一次一问,等回答再问下一个。 问题堆一起 owner 记不住,第一个还没看清就刷过,对齐效率反降。此律只约束「阻塞问 owner 的那条道」(owner Decision);Facts 自查、自裁 Decision 两条道无人类记忆瓶颈,应批量并行 / inline 轻检查点(多路 Grep/MCP/检索同时打;自裁结论直接声明),串行阻塞只是纯延迟税 + 仪式感。
- 每问先给我的推荐答案 + 理由 + 证据来源。 不是 ceremonial「你觉得呢」,是带着判断逼问——有 taste 就先出。
- 三类分道:Facts 自查 · 自裁 Decision 自决(轻检查点)· owner Decision 才阻塞问。
- Facts(代码/git 能证实:现有实现/端点归属/模式)→ 自查标 [已查证],批量并行。
- 自裁 Decision(技术选型,我有决定性证据:架构形状/接缝/字段/施工序)→ 自己拍,inline 声明「我取 X,因 Y + 证据」当轻检查点,不做阻塞提问;owner 事后 review 决策记录一次推翻,比逐个阻塞省。
- owner Decision(真需 owner 判断:业务方向/领域红线/风险偏好/两个技术上打平的方案选哪/我明说『拍不动』)→ 才停下一次一问。 判据:『owner 的答案会不会和我的证据推荐不一样?需不需要我没有的判断(业务/风险/偏好/红线)?』否 → 自裁;是 → 问。把自裁得了的 Decision 做成阻塞提问 = 仪式感,本 skill 要防的正是它——反例证据:owner 驳回浅推荐的最佳案例往往发生在普通推荐流程、非 grill,故 asking 不独占该价值,默认自裁 + 轻检查点。
- 沿决策树走,先解依赖再解叶子。 上游决策(数据模型/状态机/边界)没定之前不问下游(字段命名/UI token)。
- 对标外部实现,逼问「为何偏离」。 有外部标杆(同类 repo/论文/框架最佳实践)就 live 拉来当对抗基准(runtime 原生 web /
dag_research检索版 /context7MCP):「标准做法是 X,我们做 Y——偏离是 first-principles 的选择还是无知?理由站得住吗?」first principles > stackoverflow,但偏离要能自证,不是没看过别人怎么做就拍脑袋。
宽解岔口 → 就地开 council
grill 是纵向掘深:HITL 交互、串行、单视角,把一条决策线盘到底。遇到岔口是宽解空间 + 拿不准(多个合理方案、领域红线、架构选型),别自己拍平均答案 → 就地 fire /omd-council(dag_research council 模式):多 persona 并行出方案 + judge 择优 + 嫁接亚军亮点 → 把冠军作为「我的推荐答案」带回 grill 继续逼问。
grill 沿决策树走
├─ 岔口是 Facts / 窄解 ── grill 自己拍(先给推荐答案 + 证据)
└─ 岔口是宽解 + 拿不准 ── /omd-council 多 lens 并行 → judge 冠军 + 嫁接亚军
→ 冠军回填「我的推荐」→ grill 继续逼问 → owner 裁
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.
- 5d ago First seen · 79 lines · 116 tokens per session scan A 4bca64ef5a2f
grill is a skill published in the GitHub repository AbyssCN/oh-my-dag (43 stars, last pushed 5d ago), licensed MIT. It adds 116 tokens to every session and 2,130 once invoked, about $0.0005 per session on Opus 5.5. 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 2026-10-12.
Other skills, from other repositories
release-notes
Write release notes for a tagged release — a business-impact summary of what changed for the user, plus a thanks section naming every contributor whose commits are in the release. Short by construction.
review-this-branch
Run the nohuman review gate (fresh-session adversarial reviewer + tamper guard) over the current branch or a GitHub pull request, with no server, no database, and no onboarding, and relay the pass/fail checklist.
file-a-task
File work into nohuman (taskadd) and check on it (taskstatus) via the nohuman MCP bridge, instead of doing the work inline.
design-skill
Use BEFORE building or restyling anything a user will see — web/mobile UIs, landing pages, dashboards, components, emails, AND creative work like Three.js/WebGL/canvas/SVG scenes, data-viz, games, and animations — to make it look intentional and polished, not generic; covers tokens, type, color, layout, motion…
skill-test
Verifies that the skill system is working correctly.
interview-and-create-plan
Before starting a Ralph RLM loop, interview the user to sharpen the goal, stress-test the design, and write an authored PLAN.md (goal, definition of done, milestones, open questions, invariants) that the loop can follow. Use when a user wants to plan, scope, or stress-test work before delegating it to the supervisor /…