skill-scaffolder 精简技能架构
v1.0.0生成结构精简的 Claude Code skill 骨架,强制 routing-signal 三要素 + 五类 body taxonomy + Faithfulness gate。用户说起草/设计/新建/scaffold skill 时触发。
Security Scan
OpenClaw
Benign
high confidencePurpose & Capability
Name and description claim to scaffold compact skill skeletons; the skill is instruction-only and asks for no environment variables, binaries, or credentials — this matches a design-time scaffolding tool. Mentions of third‑party tools (e.g., pdfplumber, tabula‑py, tiktoken) appear as examples or optional help, not as required credentials or hidden capabilities.
Instruction Scope
SKILL.md and supporting files constrain the agent to produce and split skill text files, perform taxonomy labeling, run two internal gates (retire / faithfulness), and write scaffold output to a user-chosen path. There are explicit constraints to avoid hiding operational info in background/examples and to avoid overwriting existing directories. The instructions do require file-writing to the specified path, which is appropriate for a scaffolder.
Install Mechanism
No install spec is embedded (instruction-only), which is lowest risk. README offers optional install commands (git clone, npx, clawhub) that would pull code from external sources — those are external suggestions, not automated installs. If you follow those commands, verify the remote source before running them.
Credentials
The skill declares no env vars, credentials, or config paths. Some example/template text references optional tooling (tiktoken token counting, pdfplumber/tabula) but these are illustrative and not required by the skill. No disproportionate secret access is requested.
Persistence & Privilege
always:false and no install-time persistence is requested. The skill instructs writing files within the user-specified path and to avoid overwriting existing directories; it does not request to modify other skills or system-wide agent settings. Autonomous invocation is allowed by default (platform normal) but not unusual here.
Assessment
This skill is internally coherent and appears safe to use as an instruction-only scaffolder. Before installing or running: 1) If you plan to use the README's git / npx / clawhub commands, inspect the remote repository first (don't blindly run installs). 2) Be aware the agent will write files to whatever <path>/<skill-name>/ you provide—choose a safe directory and back up anything important; the scaffolder does include a non-overwrite rule but review outputs. 3) Optional examples mention tools like tiktoken, pdfplumber, tabula‑py — those are illustrative; only install them if you need the associated functionality. 4) If you allow the agent to run autonomously in environments with sensitive files, limit its file-system permissions; the skill itself does not request secrets or elevated privileges. Overall: coherent and appropriate for its stated purpose.Like a lobster shell, security has layers — review code before you run it.
latest
SkillScaffolder
Trigger
/skill-scaffolder- 用户要求起草 / 设计 / scaffold / 新建一个 skill
- 用户有 idea 要落成 skill
7 步流程(一次性产出,非引导式)
- Retirement check(Gate 0):给 3 个"典型触发场景"。若裸模型 / 现有 skill / CLAUDE.md 都能搞定 → 建议退休,停止
- Routing signal 三要素:primary capability(10–20 tok)+ trigger condition(10–20 tok)+ unique identifier(10–20 tok)
- Description draft:三要素合成 ≤120 tok;DDMIN 自检(能否再删一词仍唯一触发)
- Body taxonomy 预标:逐条规则打五类标签(
core_rule/background/example/template/redundant);拿不准读background.md - File split 布局:SKILL.md 只放
core_rule;非空拆出文件必有when:+topics:frontmatter;< 50 tok 的拆出文件合回 SKILL.md - 四反模式检查:
- ❌ examples-as-specification(规则藏在示例里)
- ❌ background 藏数字 / 阈值 / 路径 / API / env
- ❌ description 堆触发词枚举
- ❌ redundant(body 重复 refs)
- Faithfulness 自检(Gate 1):列 operational concept(动词 / 阈值 / 路径 / API / env)→ 核对都在
core_rule∪ 拆出文件;丢失则回滚到core_rule - 产出:写到
<path>/<skill-name>/:SKILL.md(core_rule only)- 非空的
background.md/examples.md/templates.md(各带 frontmatter) ARCHITECT_REPORT.md(格式读templates.md)
辅助:
- 分类拿不准 →
background.md - 对照走查 →
examples.md - 骨架 / 报告模板 →
templates.md
约束
- 不创建 < 300 tok 的 skill(Gate 0 退休)
- description ≤ 120 tok
- body
core_rule占比 ≥ 80%;不到重新归类或退休 - 不堆触发词枚举到 description
- 不把 operational 信息藏在 background / example
- 每个拆出文件必有
when:+topics:frontmatter - 不覆盖已存在的目录(若
<path>/<skill-name>/存在,改写到<skill-name>.new/)
不该用 skill-scaffolder 的情况
- 改造老 skill → 用
skill-compressor - 想要引导式创作 + eval 迭代 → 用官方
skill-creator - 只有 1–2 条规则 → 直接写进
CLAUDE.md,不建 skill
Comments
Loading comments...
