Manuscript Scaffold

Manuscript scaffold

Draft manuscripts the way a senior author would: decide what the paper claims, report only what supports those claims, and write it in plain declarative prose. This skill exists because the default model behaviour produces two things no journal would print — hook-style subheadings, and a results section that narrates the analysis history (what was tried, what came before, what was ruled out) instead of the findings. Both come from treating the work as the story. The story is the claims.

The skill is also built for scaffolding. The user often starts a manuscript before the data are complete: they know what a section will say but not the numbers. That is a legitimate working mode. Write the intended sentence with a marked placeholder rather than inventing a value or skipping the sentence. Never fabricate a number, a p-value, a citation, or a result to make a paragraph look finished.

Workflow

Run these steps in order. For a single-section request, still do steps 1–2 briefly (they take one short exchange or one paragraph of reasoning), then draft only the section asked for.

1. Intake

Establish, from the conversation, uploaded files, or by asking once and briefly:

• Target journal or format (drives length, abstract type, heading style, tense conventions). If unknown, default to the conventions in references/section-guide.md for a Nature Medicine / Cancer Discovery–type research article.
• The claims. What will the abstract assert? Usually 2–4 claims. If the user hasn't stated them, propose them from the material and confirm.
• What exists vs. what's pending. Which analyses are done, which figures exist, which data are still coming. This determines where TODO markers go.
• Exemplars. If assets/exemplars/ contains files, read them before drafting and match their register, paragraph length, and hedging style. Prefer the user's own prior papers over any generic notion of academic style.

Don't run a long interview. One message with at most three questions, or proceed on stated assumptions and list them at the top of the draft.

2. Claim ledger

Before drafting any prose, write a short claim ledger (it goes in your reasoning or in a collapsible note at the top of the draft, not in the manuscript body):

Claim 1: <one sentence, as it will appear in the abstract>
  Supported by: Fig 1a–c, Table 1, ED Fig 2
  Results subsection: "<heading>"
  Status: complete | partial (missing X) | planned
Claim 2: ...

Every results subsection must trace to a claim. Every figure panel must trace to a subsection. Anything in the source material that traces to nothing — exploratory analyses, earlier model versions, sanity checks, negative controls that don't qualify a claim — goes to the supplement, to Methods (if it's a validation step), or is dropped. The exception is a negative result the reader needs in order to trust a positive one (e.g. "the association was not explained by stage or histology"); that belongs in the same paragraph as the claim it protects, in one sentence.

The ledger is also what makes scaffolding work: a planned subsection with Status: planned gets a stub paragraph and TODO markers, not silence.

3. Outline

Produce a one-line-per-subsection outline with the claim it serves and its status. For a full manuscript this is the skeleton the user will iterate on for weeks, so get the structure right here and confirm it before writing prose. For a single section, skip confirmation and move on.

4. Draft

Draft section by section, in full paragraphs. Read references/section-guide.md for what each section contains, its length norms, tense, and the order of information in a results paragraph. Read references/headings.md before writing any heading.

Core prose rules (the full audit list with before/after examples is in references/prose-audit.md):

• Results paragraphs open with the finding, not with what was done. "NKX2-1-low tumours had shorter overall survival (HR 1.9, 95% CI 1.3–2.8)" — not "To investigate whether NKX2-1 expression was associated with survival, we performed…". Method context comes second, in a clause, or is left to Methods.
• No narration of the analysis journey. "We first tried X, which failed, so we then…" is a lab-meeting sentence. Report the analysis that stands. If a rejected approach matters for the reader's trust, one clause in Methods or a supplementary note.
• No signposting adverbs (Importantly, Notably, Strikingly, Interestingly, Remarkably, Crucially). The sentence should carry its own weight. Delete them; if the sentence collapses without one, the sentence was empty.
• Hedge with verbs and evidence, not with intensifiers. "suggests", "is consistent with", "we did not observe" — not "may potentially", "robustly demonstrates", "clearly shows".
• Numbers with uncertainty. Effect sizes carry a CI or SD; p-values accompany an effect size, never replace it. Percentages carry n.
• One interpretive sentence per results paragraph at most. Interpretation lives in the Discussion.
• No bullets, no bold, no colon-hooks, no em-dash chains, no rhetorical questions anywhere in the manuscript body.
• No rule-of-three lists for rhythm. Enumerate what there is; if there are four, list four.
• Vocabulary. Avoid the words that mark machine-written prose in this genre: landscape, paradigm, pivotal, crucial, holistic, comprehensive (as a self-description), unprecedented, novel (as a self-description), delve, underscore, highlight (as a verb for findings), leverage, harness, robust (unless it is a statistical property being tested), shed light on, pave the way, in the era of. Use the plain word.
• Don't open consecutive paragraphs the same way ("We next…", "We then…", "To address this…").
• Discussion does not recapitulate Results. One short orienting sentence at the top, then interpretation, comparison with prior work, limitations, and implication. Limitations are specific and honest, not a ritual list.

5. TODO markers (scaffold mode)

When something is missing, mark it in-line with a bracketed, typed marker so the shape of the sentence is preserved and the gap is searchable:

Rules:

• A planned subsection gets a heading and a stub paragraph in the intended register ("[TODO-DATA: methylation clocks in ASCENT] We will report whether epigenetic age acceleration differs by lineage axis position; expected direction is …"), so the outline reads as a paper, not as a list of gaps.
• Keep a TODO ledger at the end of the document: every marker, grouped by type, with its section. Update it on every revision. Resolved items are deleted, not struck through.
• Don't pad around a gap with generic prose to make the section look longer. A short stub is better than a paragraph of filler.
• Speculative framing that depends on pending data is written in the conditional and flagged with [TODO-CHECK], not asserted.

6. Prose audit

After drafting, read references/prose-audit.md and do a second pass over your own text against it. Fix what you find. Then ask yourself once: "Which sentences here would a co-author strike as filler, hedging theatre, or lab-meeting narration?" and cut them. This second pass is not optional; the first draft reliably contains the patterns the skill exists to remove.

7. Output

Deliver as a .md file (or .docx if the user asks or a submission-format deliverable is implied — then follow the docx skill). Structure:

1. A short header block: working title, target journal, date, and a 3–6 line changelog if revising an existing draft.
2. The manuscript.
3. The TODO ledger.
4. (Optional, when useful) a "Supplement / cut" section listing analyses that were deliberately moved out of the main text, so the user can see what was excluded and disagree.

Don't summarise the manuscript back to the user in chat. One or two sentences on what was drafted and the biggest open decision is enough.

Revising an existing draft

When the user brings a draft back (new data, reviewer comments, a section to fill in):

• Change what was asked. Don't silently rewrite untouched sections to your taste; if you see a problem elsewhere, note it in the changelog as a suggestion.
• Preserve the author's voice in sections they wrote. Match sentence length and hedging style. Apply the prose audit to your additions, not to their paragraphs unless asked.
• Resolve TODOs precisely. When a [TODO-NUMBER] is filled, remove the marker and the ledger entry. When a [TODO-DATA] stub is replaced with real results, rewrite the paragraph from the finding outward rather than editing the stub in place.
• Re-check the claim ledger. New data sometimes changes what the paper claims. If so, say so explicitly before editing, because that changes the abstract and the title.
• Keep a changelog at the top: what changed, what's still open, anything you'd flag.

What this skill does not do

It does not perform literature search or insert references it hasn't been given; it marks where citations are needed. It does not generate figures. It does not decide the science — where a framing choice is the author's, it flags a [TODO-DECISION] and drafts one option.