writing-artifacts

A production line: raw thoughts in, publishable artifact out. You orchestrate stations; each station's method lives in its own SKILL.md — read it at that stage, never work from memory of it.

Stage 0 — frame (one question round, then move)

Pin three things before writing anything: the SUBJECT (one sentence), the READER (who must understand it), and the OUTPUT FORMAT. Default format by destination: quick share → md · something the user will look at twice or show someone → html page or Artifact · visual metaphor / hero image needed → add an image stage. If the user already said the format, don't ask.

Reader preset — ELI5: "explain like I'm five", "dead simple", "完全外行" pins READER = knows nothing about the domain and FORMAT = Artifact carried by big pictures and few words. Simplify the language, never the facts: no jargon without a grounded term, no "simply/just/obviously".

Genre branch, decided here: in-repo SOFTWARE documentation (README, API docs, tutorials, how-to guides — anything living under the repo) swaps the structure station: run documentation-writing (Eight Rules + Diataxis, docs/ location and linking rules) as Stage 2 instead of the writing-beats/shape pair, keeping Stage 1 mining and Stage 3 stop-slop discipline unchanged. Everything else — problem writeups, incidents, decisions, long-form prose — takes the default line below.

Stage 1 — mine (diverge)

writing-fragments: dump everything known about the subject as raw fragments — evidence, timeline, feelings, half-thoughts — no structure yet. For a genuinely fuzzy idea (not a known incident), run adhd first for parallel divergence, then fragment the survivors.

Stage 2 — structure (converge)

Pick ONE:

• writing-beats — when the reader must be LED somewhere: problem

narratives, incident writeups, decision rationales. Grounds each term before a beat leans on it.

• writing-shape — when the material is already roughly ordered and just

needs shaping paragraph by paragraph.

Stage 3 — tighten

edit-article over the draft, with stop-slop loaded as criteria for the whole run (not just this stage): no filler, no hedging, no AI-slop cadence.

Stage 4 — render (format branch)

• md / plain prose → deliver the tightened text directly; done.
• html page / Artifact / diagram / image → hand the FINISHED TEXT to

using-design-skills as the content contract and let IT compose the pipeline (it owns authority selection, imagegen delegation, and the screenshot-evidence quality loop). Never style inline yourself; the text is frozen content by this point — design changes layout, not wording.

Stage 5 — report the edit (in the turn, not in the artifact)

After delivering, add at most five lines naming what changed and why. One line per edit class, not per sentence: the jargon you grounded, the assumption you made explicit, the structure you reordered, what you cut. Name the problem, not the prettier wording.

When the input was EXISTING text, each line is before → after, quoting the shortest fragment that carries the problem. When the text is new, drop the before half and report what you cut and why.

This report never enters the artifact.

Boundaries

• Words first, pixels second: never enter Stage 4 with an untightened draft —

redesigning slop produces well-dressed slop.

• One pass back is allowed: if rendering reveals a structural hole (a section

that can't be visualized because it was never actually explained), return to Stage 2 for THAT section only, then re-render.

• This skill owns written deliverables about a subject — prose AND in-repo

software docs (the genre branch in Stage 0 picks the structure station). A plan page from a discussion → html-plan directly.