author-stepwise-skill

This skill scaffolds a multistep skill. A docs-only context-mill skill made of an ordered chain of reference files. The agent reads one file at a time and never sees later files until it finishes the current one.

When this shape is right

This shape gives you three things:

• Bounded context per step. The agent reads one focused file at a time. Token use stays small. Attention stays on the current goal. The shape works the same for two steps or twenty.
• A clean checkpoint per step. Each step finishes when its verifiable objective completes. Progress is observable. You can pause, check the run is on track, and resume from the next file.
• Constrained scope. The agent acts only on what it has been shown. It does not preload future work, branch, or chase tangents.

Use the multistep pattern when all of these are true:

• The work is sequential. Step N depends on step N-1's output.
• Each step has a narrow job. If you can describe step N in one paragraph, it is narrow enough.
• The skill is docs-only. There is no example project to bundle.

If the skill is just a bag of related references that the agent can read in any order, do not use this shape. Use a flat docs-only skill. Extra ceremony hurts.

The mechanics that follow (references.preamble, per-file next_step, "do not Glob/ls/find", "do not preload", "do not re-read earlier files") make progressive disclosure the path of least resistance. Do not weaken them. A step that says "for context, also read X" defeats the pattern.