M
Michael-Preibisch/mpreibisch-skills
My custom skills for Claude and Codex
★ 1·更新 2026-11-17
在 GitHub 查看 ↗直链 · 无中转页
- ★ 数取值
- 1 ★ · Michael-Preibisch/mpreibisch-skills · 取自 GitHub 搜索 API · (2026-11-19 20:00:48 核对)
- 来源明细
- 从该仓库的 GitHub 页面直接读取 · 条目类型:开源项目
- 收录与活跃
- 收录于 2026-11-18 · 最近见到 2026-11-20 · 共见到 3 次 · 被 1 个来源列出
- 评分构成
- 综合分 103 · 1 ★ · 1 个来源 · repo
常见问题(答案都来自我们的记录)
它适合谁 / 用来做什么?
我们把它归在这些标签下: 开源项目。标签来自条目自身,不是我们强加的分类。
谁维护它、还在更新吗?
仓库
Michael-Preibisch/mpreibisch-skills;我们收录于 2026-11-18;最近见到 2026-11-20。这里没有"官方认证"或"质量评级"——我们只记录来源与时间。README 原文(英文,未翻译) · 在 GitHub 读完整 README →
# Personal cross-model skills
以下摘录自该仓库的 README(片段)
Two installable plugins pair Codex and Claude Code through their CLIs. Each includes
consensus, adversarial-review, brainstorm, delegate-explore, delegate-implement,
the pinned clear-writing skill, artifact-standards, pr-review-ask, and decisions.
| Calling host | Plugin | Receiving CLI | Installation |
| Codex | mpreibisch-codex | claude | Codex README |
| Claude Code | mpreibisch-claude | codex | Claude Code README |
decisions calls the OpenAI Decisions API for fixed judgments over supplied text.
Both hosts can use it proactively for yes/no questions, labels, or fixed scores.
It can gate an expensive polling or triage agent, with every uncertain result or
error falling back to the LLM. It needs Python and an API key, without a peer CLI.
Presets cover progress watching, review triage, failure retries, model-tier routing, and relevance filtering.
All bundled presets default to shadow mode. Overridden preset rules also default to shadow, even if the source is calibrated.
The decision shape also covers custom questions and one-off categorization.
The authoring guide defines evidence, avoided operations, and fallbacks.
External templates show caller-defined request routing and severity rubrics.
Use --preset-file PATH for portable configuration, --shadow for any gate, and
report --log FILE --labels FILE for offline evaluation against independent outcome IDs.
The binary report evaluates one defined skip decision; it does not validate routing labels or tier adequacy.
Loop acknowledgements and scheduled full checks remain caller-owned.
pr-review-ask drafts a request to review one or more GitHub PRs in the user's own
voice: motivation first, one line per PR, confirmed stacks kept visible, linked issues
from GitHub. It fetches PR facts with gh and shows the draft in chat. If the user
names a destination and the host has a draft tool for it, it saves a draft there. It
never sends.
brainstorm asks the other model for ideas without showing it yours, then swaps the
lists so each side builds on and challenges the other's, and returns a ranked shortlist
that keeps the dissent. It never forces agreement; consensus does that.
artifact-standards starts no cross-model call. It holds the
house standards for every shareable HTML page: page shape, palette, a light-first
theme contract with an in-page toggle, the writing gates, terms-first for explainers,
source links, and a pre-publish checklist. It ships a page shell and design guidance
inside the skill, so it needs no host design skill. Delivery adapts to the host: the
agent always writes one self-contained HTML file outside every repository worktree,
then publishes it if the host has a publisher and the user authorized publishing, or
hands back the local file if it does not.
Every cross-model request carries the broader goal, motivations, constraints,
decision rationale, intended outcome, scope, and expected output. The runner rejects
missing context fields. The main agent supplies meaningful context and verifies the
returned evidence. Only delegate-implement grants scoped implementation writes.
The handoff forbids external side effects, and the helper rejects requests granting
consensus, adversarial-review, brainstorm, delegate-explore, delegate-implement,
the pinned clear-writing skill, artifact-standards, pr-review-ask, and decisions.
| Calling host | Plugin | Receiving CLI | Installation |
| Codex | mpreibisch-codex | claude | Codex README |
| Claude Code | mpreibisch-claude | codex | Claude Code README |
decisions calls the OpenAI Decisions API for fixed judgments over supplied text.
Both hosts can use it proactively for yes/no questions, labels, or fixed scores.
It can gate an expensive polling or triage agent, with every uncertain result or
error falling back to the LLM. It needs Python and an API key, without a peer CLI.
Presets cover progress watching, review triage, failure retries, model-tier routing, and relevance filtering.
All bundled presets default to shadow mode. Overridden preset rules also default to shadow, even if the source is calibrated.
The decision shape also covers custom questions and one-off categorization.
The authoring guide defines evidence, avoided operations, and fallbacks.
External templates show caller-defined request routing and severity rubrics.
Use --preset-file PATH for portable configuration, --shadow for any gate, and
report --log FILE --labels FILE for offline evaluation against independent outcome IDs.
The binary report evaluates one defined skip decision; it does not validate routing labels or tier adequacy.
Loop acknowledgements and scheduled full checks remain caller-owned.
pr-review-ask drafts a request to review one or more GitHub PRs in the user's own
voice: motivation first, one line per PR, confirmed stacks kept visible, linked issues
from GitHub. It fetches PR facts with gh and shows the draft in chat. If the user
names a destination and the host has a draft tool for it, it saves a draft there. It
never sends.
brainstorm asks the other model for ideas without showing it yours, then swaps the
lists so each side builds on and challenges the other's, and returns a ranked shortlist
that keeps the dissent. It never forces agreement; consensus does that.
artifact-standards starts no cross-model call. It holds the
house standards for every shareable HTML page: page shape, palette, a light-first
theme contract with an in-page toggle, the writing gates, terms-first for explainers,
source links, and a pre-publish checklist. It ships a page shell and design guidance
inside the skill, so it needs no host design skill. Delivery adapts to the host: the
agent always writes one self-contained HTML file outside every repository worktree,
then publishes it if the host has a publisher and the user authorized publishing, or
hands back the local file if it does not.
Every cross-model request carries the broader goal, motivations, constraints,
decision rationale, intended outcome, scope, and expected output. The runner rejects
missing context fields. The main agent supplies meaningful context and verifies the
returned evidence. Only delegate-implement grants scoped implementation writes.
The handoff forbids external side effects, and the helper rejects requests granting
收录徽章(可选)
想把收录标记放进 README 就复制这行,可选,不影响收录。
[](https://anclum.com/zh/s/michael-preibisch-mpreibisch-skills/)
- 来源
- github · https://github.com/Michael-Preibisch/mpreibisch-skills
- 按领域浏览
- 开源项目
- Traction
- 1 ★
- 收录于
- 今天
更多条目
所属领域与标签
每个标签都能点进去看同类的全部条目。数字是当前库里的真实条数。