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.
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 →
· something the user will look at twice or
show someone →
page or Artifact · visual metaphor / hero image needed →
add an image stage. If the user already said the format, don't ask.
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
(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
discipline unchanged.
Everything else — problem writeups, incidents, decisions, long-form prose —
takes the default line below.
: 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
first for
parallel divergence, then fragment the survivors.
over the draft, with
loaded as criteria for the
whole run (not just this stage): no filler, no hedging, no AI-slop cadence.