Manuscript Writing

Manuscript-Writing — write and organize in Khan's voice

This skill helps draft, organize, and revise the user's (Muhammad Ali Khan's) manuscripts and abstracts so they read as unmistakably his and are structured the way he structures them. It was built by analyzing ~23 of his real submitted manuscripts and abstracts spanning clinical/epidemiology cohort work, AI/LLM-methods papers, and systematic reviews / meta-analyses / network meta-analyses. The goal is not generic "good academic writing" — it is his recognizable fingerprint.

When you're invoked

1. Scope in one pass (don't interrogate). Infer and state your assumptions: (a) document type — full manuscript vs conference abstract vs grant deliverable; (b) research type — clinical/epi cohort, AI/informatics-methods, or evidence synthesis (SR / MA / NMA, living/interactive?); (c) target venue if stated; (d) which section(s) the user wants. Only ask if genuinely blocked. Verify current venue section labels, word limits, and submission rules from the official source when they affect the deliverable.
2. If it's an SR / MA / NMA, use the manuscript-srma skill in the same turn for reporting mechanics (PRISMA/PRISMA-NMA, GRADE/SoF, and journal adaptation). This skill supplies voice and organization; manuscript-srma governs scientific integrity, reporting mechanics, and placeholder syntax whenever the two are composed. If it is unavailable, apply the fallback conventions in references/statistical-reporting.md and references/section-playbook.md without blocking.
3. Draft or edit using the moves below. For anything section-specific, read the matching reference file — don't work from this summary alone when you're actually writing that section.
4. Never fabricate. For every newly introduced gap, use one complete canonical token: ⟦PLACEHOLDER — concise description of the needed input⟧. List every placeholder at the end so it can be cleared. Treat (X), XX%, $XX, 202X, IRBXX-XXXX, and NOTE TO ED: already present in a legacy draft as unresolved placeholders too, but do not introduce new legacy tokens unless the user explicitly requests that house-token style.
5. Confirm authorship metadata. Never infer or insert author order, affiliations, co-senior status, or corresponding-author details from the house references. Use only the project-supplied byline or ask for confirmation when the deliverable includes authorship metadata.

Reference files — read the one you need

Progressive disclosure: this file holds the always-true core. Open the relevant reference before drafting that part.

• references/voice-and-phrasing.md — the full voice signature, sentence/grammar rules, the phrase bank (his actual connectors, pivots, closers), and 25 verbatim example sentences. Read before drafting ANY prose.
• references/section-playbook.md — how he builds each manuscript section (Title, Introduction, Methods, Results, Discussion, Conclusion, Limitations). Read before writing that section.
• references/abstracts.md — the conference-abstract and grant-abstract playbook (heading sets, two-sentence Background funnel, dense Methods paragraph, inline effect triples, the Table close). Read before writing an abstract.
• references/statistical-reporting.md — his quantitative fingerprints: the inline effect triple, the (coefficient; p) compact pair, the absolute "N per 1000" translation, heterogeneity/GRADE/NMA-ranking conventions, ML-metric reporting. Read before writing any Results or reporting any number.
• references/organization-and-workflow.md — the skeleton-first / tables-first project workflow, Methods↔Results parallelism, journal-adapter selection, versioned file naming, legend conventions, and the end-of-draft copyedit checklist. Read when organizing a project or preparing a draft for submission.

The voice in one paragraph

Khan writes in a rigorously scaffolded, methods-forward register. Every piece opens on a broad field-level pressure (evidence outpacing clinicians; a disease's burden; AI outrunning governance) and funnels, paragraph by paragraph, to one unresolved gap, then pivots on a fixed hinge — "To address this gap, we..." — into a first-person-plural statement of what the work does. He owns the work in the active voice (we developed, we found, our approach) and uses agentless passive only for procedures (Data were abstracted). He defines every abbreviation on first use and reuses it relentlessly, often coining a Title-Case branded framework (LIvE, LISR, ONCAIR) or a compact study nickname (dFLOT, aIO) and reusing it across papers. Two temperamental signatures run through everything: constructive humility (limitations are quarantined and each is paired, in the same breath, with a mitigation or future-work rebuttal, so they read as a roadmap; work closes on a standing caveat — prospective/multi-site validation before deployment; NMA rankings complement, not replace, RCTs) and ambition (first-in-class, paradigm shift from reactive to proactive, an expansive vision sentence to close). He is reproducibility-minded: exact software versions, PMIDs, cohort counts, model temperature/zero-shot settings, search cutoff dates, reviewer initials.

The core moves (true in every document type)

These recur across clinical, AI, evidence-synthesis, and abstract writing — they ARE the voice. Apply them by default.

1. The problem-first funnel. Open on a broad pressure or a hard epidemiologic/prognostic statistic with a citation — never on the specific study object. Narrow over 3–4 paragraphs (manuscript) or two sentences (abstract) to the single gap. Put the study object only in the last intro paragraph. ("Around 30% of all newly diagnosed bladder cancer cases are muscle-invasive in nature…")
2. The fixed hinge. Pivot from problem to your work with "To address this gap, we…" / "To overcome this limitation, we have developed…" / "To this end, we…", then end the intro with a thesis+roadmap sentence and an explicit aim or "We hypothesized that…".
3. First-person-plural ownership. Every authorial action and claim is we/our, active voice (we performed, we found, our analysis clearly indicates). Never first-person singular. Passive only for procedures. Contrast "our approach" against "current approaches / existing frameworks" to stake the novelty claim.
4. Define-and-brand. Spell out every abbreviation in parentheses on first use, then use the short form. Where warranted, coin a memorable Title-Case, parenthetically-acronymed framework and reuse the brand across papers; for AI systems, name the modules (The Watcher, The Scanner, The Extractor).
5. Explicit enumeration. Enumerate open questions, implications, strengths, and limitations with Firstly/Secondly/Thirdly/Lastly or First/Second/Third.
6. Results describe; interpretation lives only in the Discussion. The Results section is a pure, descriptive transcription of the exhibits — report what was measured, in whom, and the value with its exhibit pointer, then stop. No interpretation in Results: no mechanism, no "this suggests / likely reflects / because / consistent with / indicating," no clinical significance, no comparison to prior work. All of that is quarantined to the Discussion. Write Methods as an ordered pipeline of short labeled subsections and title each Results subsection identically, so a variable tracks from method to result by name; Results always opens with the total N/denominator. (If a Results sentence would still be true with the numbers removed, it's interpretation — move it to the Discussion.)
7. Template-and-swap parallelism, then tether to an exhibit. When the same measure recurs across subgroups, arms, or subtypes, write the sentence once and reuse it verbatim for each stratum — change only the stratum name and its numbers; keep the sentence structure, verb, clause order, and punctuation identical (the PV/ET/MF order is recycled word-for-word across sentence, table, and figure, so strata line up column-wise). Report the overall population first, then the strata in the same fixed order every time. End every quantitative claim by pointing to a specifically numbered exhibit (Table 2; Supplement Figure S11); Results cite exhibits, never the bibliography, with detail pushed to a heavily indexed supplement.
8. Effect-estimate triple + justified absolute translation. Report estimates inline as (HR: 0.72; 95% CI: 0.56–0.94; I²: 87%; Supplement Figure S11). Add "N fewer/more events per 1000 patients" with its own CI only when the effect measure, baseline risk or survival, time horizon, and transformation support that calculation; otherwise state what input is missing. (See references/statistical-reporting.md.)
9. Candid limitations with supported mitigations. Quarantine limitations into their own block, opened with "However, the findings of this study should be interpreted in the context of some limitations." Pair a weakness with a mitigation or future-work step only when the project evidence supports it and it genuinely addresses the concern; otherwise state the residual limitation plainly.
10. The reflexive close. Land on the standing caveat (prospective/multi-site validation warranted before deployment; NMA rankings complement not replace RCTs), then an ambitious impact/utility sentence.

Connectors and tics — use, but don't autopilot

His fixed connector set opens sentences and paragraphs: However, Moreover, Furthermore, In addition/Additionally, Notably, Importantly, Interestingly, Conversely, Taken together, Of note, It should be noted that. Use them — they're his voice — but don't start several consecutive sentences with "However," / "In addition,"; that tips from characteristic into mechanical. The full phrase bank (pivots, novelty claims, the templated results verb-frame, the standing caveats, the limitations opener+rebuttal, the strengths formula) is in references/voice-and-phrasing.md.

Watch his recurring slips (fix, don't reproduce)

His voice is worth imitating; his copyedit slips are not. When editing or before "finishing," hunt these — they survive many of his versions unless deliberately targeted:

• Doubled words ("ATM mutations ATM mutations"), run-together words ("performednetwork", "withing"), dropped articles, subject–verb disagreement, double negatives.
• Recurring misspellings: Certainity → Certainty, continuosly → continuously.
• Duplicated pasted paragraphs or figure legends; mixed Fig. vs Figure cross-references.
• Inconsistent CI delimiters (semicolon vs comma) — standardize on semicolons.
• Mixed dispersion notation in one paper (mean ± SD vs mean (SD = )) — pick one per manuscript.
• Left-in draft scaffolding: (X) citation placeholders, XX%/$XX/202X, IRBXX-XXXX, NOTE TO ED — clear all before submission and verify every reference.

Editing vs drafting

• Drafting a section: read the matching reference file, then write it to the architecture and voice there. Match section labels and citation style to the target venue up front (decide once; keep the whole draft consistent).
• Editing his draft toward his voice: preserve his content and claims; restructure toward the funnel/parallelism, tighten runaway sentences, enforce the effect-triple plus any justified absolute translation, pair limitations only with supported mitigations, and run the copyedit-slip hunt. Don't sand off the voice — keep the enumeration, the branded terms, the ambition. The failure mode is over-neutralizing him into generic journal prose.
• Organizing a project: follow references/organization-and-workflow.md — lay a reserved-space skeleton, build tables/exhibits first, write prose against them, derive the abstract from the finished manuscript last.