一键在 Manus 中运行任何 Skill

skill-forge

星标4

分支0

更新时间2026年6月9日 15:45

Crystallize a captured workflow journal into a validated Claude skill. Use this whenever the user says "turn this into a skill", "make this repeatable", "build a skill from this", or when a workflow-capture worklog is ready to become a skill. Also use to improve or extend an existing produced skill from a new worklog.

安装

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。

在 Manus 中运行

来源

Mfrostbutter

Mfrostbutter/skill-forge

打开 GitHub 仓库查看创作者相关仓库

下载

在 Manus 中运行

Skill Forge

Turn a WORKLOG.md into a skill. The journal's tags map directly onto skill sections, so synthesis is mostly routing.

Step 1: Read the worklog

Read the full _worklogs/<date-slug>/WORKLOG.md. If the user references a workflow without a path, find the matching worklog.

Step 2: Route by tag

DECISION + SEQUENCE -> the procedure: ordered, imperative steps.
CORRECTION + DEADEND -> guardrails: "do not", "always", warnings, each with the one-line reason.
CONVENTION + DISCOVERY -> reference files and compatibility notes.

A DEADEND that blocked progress usually becomes an early SEQUENCE step in the procedure so it cannot recur.

Step 3: Draft SKILL.md

Read references/skill-anatomy.md first. Then produce:

name: kebab-case identifier.
description: the trigger. Make it pushy and list concrete contexts; Claude tends to undertrigger skills. Put all "when to use" info here, not in the body.
body: imperative voice, under ~500 lines. Explain why a rule matters rather than stacking bare MUSTs.
references/: push variant-specific or long details out of the body. Files over ~300 lines get a table of contents.
scripts/ and assets/ only if the workflow has deterministic steps or output templates.

Step 4: Validate

Adapt to the environment.

Subagents available (e.g. Claude Code): run 2 to 3 realistic test prompts, baseline vs with-skill, grade, review the results.
No subagents (e.g. a chat client): run test prompts inline, one at a time, using the skill yourself. Skip quantitative benchmarking; rely on qualitative review.

Iterate: run -> review -> revise SKILL.md -> rerun. Optimize the description for triggering once stable.

Step 5: Archive and link

Move the source worklog to skills/<name>/.source/WORKLOG.md for provenance. Archive, never delete. Symlink or copy the produced skill into the runtime skill path so it loads.

Step 6: Package (if available)

If a packaging script is available, package the skill and hand back the artifact path.

同仓库更多 Skills

同仓库

figma-design

Mfrostbutter/skill-forge

Design, build, or edit a Figma file through a design connector (plugin-API calls verified with screenshots). Use whenever the task is to create or change a Figma design, mock up a UI, logo, infographic, or design system, build styles or components, or hand a design to dev. Triggers even if the user does not name Figma but asks to "mock up", "design a screen", or "build a component library". Encodes the hard rule that page structure and layer naming come before any frame.

2026-06-094

workflow-capture

Mfrostbutter/skill-forge

Maintain a durable, tagged journal of a complex workflow so it can later be distilled into a reusable skill. Use this whenever beginning a complex, multi-step workflow (design work, automation build, CAD, RAG pipeline, website work, or anything non-trivial you might want to repeat), even if the user does not say "capture" or "log". If the work has sequencing, corrections, or hard-won discoveries worth keeping, start a worklog. Also use on resume to read an existing worklog after a context reset.

2026-06-094

name	skill-forge
description	Crystallize a captured workflow journal into a validated Claude skill. Use this whenever the user says "turn this into a skill", "make this repeatable", "build a skill from this", or when a workflow-capture worklog is ready to become a skill. Also use to improve or extend an existing produced skill from a new worklog.

Skill Forge

Turn a WORKLOG.md into a skill. The journal's tags map directly onto skill sections, so synthesis is mostly routing.

Step 1: Read the worklog

Read the full _worklogs/<date-slug>/WORKLOG.md. If the user references a workflow without a path, find the matching worklog.

Step 2: Route by tag

DECISION + SEQUENCE -> the procedure: ordered, imperative steps.
CORRECTION + DEADEND -> guardrails: "do not", "always", warnings, each with the one-line reason.
CONVENTION + DISCOVERY -> reference files and compatibility notes.

A DEADEND that blocked progress usually becomes an early SEQUENCE step in the procedure so it cannot recur.

Step 3: Draft SKILL.md

Read references/skill-anatomy.md first. Then produce:

name: kebab-case identifier.
description: the trigger. Make it pushy and list concrete contexts; Claude tends to undertrigger skills. Put all "when to use" info here, not in the body.
body: imperative voice, under ~500 lines. Explain why a rule matters rather than stacking bare MUSTs.
references/: push variant-specific or long details out of the body. Files over ~300 lines get a table of contents.
scripts/ and assets/ only if the workflow has deterministic steps or output templates.

Step 4: Validate

Adapt to the environment.

Subagents available (e.g. Claude Code): run 2 to 3 realistic test prompts, baseline vs with-skill, grade, review the results.
No subagents (e.g. a chat client): run test prompts inline, one at a time, using the skill yourself. Skip quantitative benchmarking; rely on qualitative review.

Iterate: run -> review -> revise SKILL.md -> rerun. Optimize the description for triggering once stable.

Step 5: Archive and link

Move the source worklog to skills/<name>/.source/WORKLOG.md for provenance. Archive, never delete. Symlink or copy the produced skill into the runtime skill path so it loads.

Step 6: Package (if available)

If a packaging script is available, package the skill and hand back the artifact path.