Free Product

Skill Glue

它不是工作流平台，也不是全自动大脑。它是一个低 token 成本的 skill 组合器，让你先把 skill 拆小，再按规则拼成可执行流程。

Why now

当 token 变贵，真正有价值的不是 giant skill，而是可拆、可换、可估算的组合能力。

像拼零件，不像写巨型 prompt

每个 skill 先拆成小原子节点，再决定串行、并行、替换和缓存，不靠一个大 prompt 撑全场。

先估算，再执行

先给出步骤结构和 token 预算，再决定要不要真的跑，这比先烧 token 再后悔经济得多。

给 Agent 读，也给人读

CLI 输出支持 JSON 和 Markdown，两种形态都围绕同一套命令：compose / estimate / explain。

CLI surface

第一版只做三条命令：组合、估算、解释。

compose

skill-glue compose <manifest> --format json|markdown [--sample-run]

Generate the composed workflow plan with step order, validation checks, and cacheable stages.

estimate

skill-glue estimate <manifest> --format json|markdown

Estimate total token budget and show which steps drive cost.

explain

skill-glue explain <manifest> --format json|markdown

Explain why steps are ordered this way and what can be replaced or cached.

Manifest examples

先给出标准 manifest，再决定怎么组合、替换和缓存。

serial-brief-stack

A narrow serial flow for intake -> research squeeze -> final brief generation.

3 steps · /examples/skill-glue/serial.json

parallel-content-stack

A mixed serial + parallel flow that prepares hooks and evidence separately before assembly.

4 steps · /examples/skill-glue/parallel.json

replaceable-agent-stack

A flow that can swap a heavy planner for a lighter planner depending on branch count.

4 steps · /examples/skill-glue/replaceable.json

Sample output

输出结果先服务于“看懂”和“估算”，不先假装自己是执行平台。

{
  "id": "parallel-content-stack",
  "execution_order": [
    ["topic.intake"],
    ["hook.builder", "evidence.bundle"],
    ["assembly.pass"]
  ],
  "estimated_cost": {
    "total_score": 5,
    "budget_band": "low"
  },
  "cacheable_steps": ["topic.intake", "hook.builder", "evidence.bundle"]
}

Source & proof