A Claude Code workflow for less obvious AI writing
Two turn Stefan Zweig’s editing habits into a repeatable AI writing workflow. `zweig-write` makes the first draft denser by reducing filler, hedging, repeated points, and unnecessary decoration. `zweig-refine` does not simply ask Claude to improve the text in one pass.
It reads the full file, removes any sentence, bullet, or word that can go without changing the meaning, and repeats the process up to four times if edits were made. The key detail is that the refinement uses real , so the file is actually edited instead of only being reviewed in Claude’s reasoning. A subagent then receives only the final draft and checks what is confusing or missing.
If the subagent finds a gap, the main refinement step adds back enough clarity. A separate `` skill was also made by having Claude study human-written posts and turn the recurring style choices into writing rules. The example uses a short explanation about mantis shrimp vision: a plain first draft becomes more personal, then gets tightened and clarified after the fresh-reader check.
Key points
- `zweig-write` helps duce a tighter first draft.
- `zweig-refine` edits the file directly and repeats the trimming process up to four times.
- A subagent checks only the final draft for confusing parts and missing information.
- A `` skill can be built from a person’s own past writing.
- The approach still needs human judgment about substance, clarity, and when to stop editing.