Outcome: a written plan the author can draft from, and, when a draft exists, a critique of the draft against that plan. Writing is not sequential; deciding what goes where, for whom, and why is the work. The agent does that work with the author and leaves the sentences to them unless asked otherwise.
Scope: long-horizon work only, per the repository's AGENTS.md.
Adapted from the writing sections of Carlini's essay. The principles it draws on are in the taste skill's references; this skill turns them into a plan.
Build the plan
Ask in this order, one at a time, and write the plan as answers arrive.
- The one idea, in one sentence. An "and" means two papers; settle which one before continuing.
- The reader. Default: the author six months ago. What did that person believe? What would they have needed to hear to think this idea was worth pursuing? If the paper argues against a community's belief, the reader is one member of that community, and the claim may need to arrive last (see episode E8 in the taste references).
- The best-case conclusion. What does the reader take away beyond the results? If it restates the abstract, the plan is not done.
- The introduction as a story. Where does the reader stand in paragraph one? How far is the journey to the setting where the idea makes sense: one sentence, a paragraph, or two pages? Two pages is the ceiling.
- The figure storyboard. Which figure shows the problem, which the method, which the results? Each must carry a one-sentence takeaway in its caption and be understood without the text.
- Section order with reasons. For each section, one line: what it does for this reader and why it sits here. Background exists to widen who "the reader" is; after it, assume they know what it taught.
- Abstract shape. Topic, problem, result or method, the one not yet covered, why it matters; or claim, evidence, impact for a broad topic or narrow audience. It will contain at least one specific number and no hedge the evidence does not require.
Write the plan as a short document with those seven parts. Stop there unless the author asks for prose.
Critique a draft against the plan
When a draft exists, read it only against the plan: does each section do what the plan says, for the stated reader? Report mismatches as questions, not rewrites. Check the fixed points: one idea, reader met in paragraph one, introduction within two pages, captions standalone, conclusion answers "so what", abstract has a number and no unearned hedge. Do not change claims, numbers, or the scope the evidence supports. Suggest the author read the draft aloud or through text-to-speech and mark what does not land.
Output
The seven-part plan, then, when a draft exists, the list of mismatches as questions. Nothing else.