IEEE-Style Scientific Writing

Use this skill to create or rebuild IEEE manuscript prose — to construct the argument, not only to polish finished sentences. For sentence-level language work, hand off to ieee-polishing.

Core stance

• Author evidence comes first. Never invent results, numbers, mechanisms, references,

benchmark schemes, channel models, simulation parameters, network architectures, training data, novelty claims or limitations. If essential evidence is missing, write an explicit [PLACEHOLDER: ...] or ask the user — do not fill the gap.

• Write the argument before the sentences. Decide what each paragraph must prove, then draft.
• Make the paper easy to judge on relevance, novelty, soundness, and reproducibility —

these are the axes IEEE reviewers actually score.

• Ambitious but bounded claims. Tie every claim to the evidence that supports it and stop

the claim where the evidence stops.

• IEEE register, not magazine narrative: precise, technical, US English, numbered citations

[n], equations numbered only for the skeleton of the argument.

When to open extra files

Keep this file lean. Open a reference only when its row matches the task.

For simulation content (which benchmark schemes/metrics/sweeps to run) use the ieee-experiments skill; for figures use ieee-figure; for reference formatting use ieee-citation; for a WCL/CL letter use ieee-letter.

Intake — establish before drafting

• section: title, abstract, Index Terms, introduction, related work, system model, problem

formulation, proposed solution/algorithm, performance analysis, numerical results, discussion, conclusion, or full outline

• paper type: optimization/resource-allocation design, performance analysis (closed-form

outage/BER/rate), new system/architecture (RIS, NOMA, ISAC, cell-free, ...), learning-based (DL/deep-unfolding/DRL for a PHY/MAC task), signal-processing algorithm, or measurement/empirical study

• core claim: what the paper actually demonstrates, in one sentence
• evidence: derivations/theorems, simulation curves, benchmark comparisons, complexity analysis
• boundary: where the claim stops (CSI assumption, channel model, scope, conditions)
• target: the specific journal (JSAC/TWC/TCOM = transactions; WCL/CL = letter) and its current

page/abstract rules; if the target is unclear, use references/journal-fit.md before drafting

If target, core claim, evidence, or boundary is missing, surface the gap first. You may still produce a scaffold with explicit placeholders.

Writing workflow

1. One-sentence argument: `In [problem/system], we show [advance] using [approach],

supported by [evidence], under [boundary].` Get the user to confirm it.

1. Choose section architecture (see Section defaults below).
2. Map each paragraph to one job: context, gap, prior-method limitation, contribution,

design, result, comparison, mechanism, implication, or limitation. One paragraph, one job.

1. Draft from evidence outward. Keep each claim next to the data that support it.
2. Calibrate verbs to evidence: show / demonstrate (strong, measured) →

indicate / suggest (inferential) → may / could (possible). Downgrade if evidence is thin.

1. Remove unsupported novelty and universal claims ("first ever", "optimal", "in all cases").
2. Paragraph-flow check: does the first sentence of each paragraph state its message, and

does each sentence follow from the previous one? (See references/body-revision.md.)

1. Return prose plus a claim–evidence map and a short note on assumptions / missing inputs.

Section defaults (IEEE)

Title

object/system + method/capability + task or consequence. Concrete over slogan. Avoid grant-style aims and overbroad field claims. Keep abbreviations minimal and standard.

Abstract + Index Terms

Single self-contained paragraph: ~150–250 words for transactions, usually ~75–100 words for WCL/CL letters under current ComSoc guidance, no citations, no undefined abbreviations, no equations. Default order: context/problem → gap → approach → key quantitative result → implication/boundary. Follow immediately with Index Terms (5–8 items, alphabetical, lower case except proper nouns). Details and method-paper variants: references/abstract.md.

Introduction (Section I)

application/importance → how prior methods work → their limitation + the technical reason → the unresolved gap this paper targets → our approach and why it works → contributions list → paper organization. End with an explicit contributions list ("The main contributions of this paper are summarized as follows:") and a one-sentence organization paragraph ("The remainder of this paper is organized as follows..."). Reason backward (what challenge, what contribution, why it works) before writing forward. Templates and sentence skeletons: references/introduction.md.

Related Work (Section II, optional)

topic scope → representative methods grouped by approach → limitation tied to this paper → our distinction. Group by technical theme and mechanism, never by year. Do not re-prove the gap already made in the Introduction.

System Model (Section II)

network/topology → channel model → transmit & receive signal → performance metric (as an equation). Describe the signal flow (what is sent, how the channel transforms it, what is received, why it matters), not a symbol dump. End by defining the metric the paper optimizes or analyses. Put a "Notation" paragraph (vector/matrix/set conventions) at the end of Section I or the start of Section II. Details: references/method.md. For an emerging architecture (RIS, NOMA, ISAC, near-field, movable antenna, cell-free, mmWave/hybrid, semantic), check the per-paradigm must-have elements, distinctive constraint, and pitfalls in references/system-model-paradigms.md.

Problem Formulation (Section III)

objective → every constraint, each explained → problem class → why it is hard. State which quantities are optimization variables; classify the problem (non-convex / MINLP / fractional) and name the difficulty that motivates the solution. Details: references/method.md.

Proposed Solution / Algorithm (Section IV)

reformulation/transformation → algorithm box → convergence → complexity. Make the path from the hard problem to a solvable one explicit and motivated; label each step exact vs approximate. For SCA/SDR/ADMM/fractional-programming patterns and proof structure: references/theory-and-proofs.md.

Proposed Network / Learning-based Solution (learning papers)

architecture (layers, dimensions, parameter count) → input/output and normalization → loss function tied to a comms metric → training algorithm → inference complexity & latency vs the model-based benchmark. Prefer deep unfolding (unrolled WMMSE/ADMM/SCA) over a black-box net, and justify the choice; the loss must be a communications metric (NMSE, negative sum rate), not an abstract proxy. Details: references/learning-based-comms.md.

Performance Analysis (analytical papers)

assumptions → derivation → theorem (closed-form outage/BER/rate) → asymptotic corollary (diversity order / DoF). Long proofs go to an Appendix; plan a simulation overlay to validate every derived expression. Details: references/theory-and-proofs.md.

Numerical / Simulation Results

simulation setup (parameters table) → benchmark schemes → analysis-vs-simulation validation → performance vs benchmarks → parameter sweeps (SNR / antennas / users) → convergence & complexity → robustness (imperfect CSI). Build an evidence ladder, claim-first per subsection. For what to produce, use ieee-experiments; for narrative, references/results-discussion-conclusion.md.

Discussion (if separate)

central advance → what the curves mean → relation to prior work → constraints (CSI / scope) → outlook. Interpretation and limitations live here; do not replay results figure by figure.

Conclusion

contribution → decisive evidence → implication → boundary/future work. No new data, no unsupported promises, no re-derivation of equations.

Output format

1. Draft: the requested prose (or LaTeX if the user is in IEEEtran).
2. Section outline: 3–7 compact bullets when drafting a whole section.
3. Claim–evidence map: Claim … | Evidence … | Status: supported / needs evidence.
4. Assumptions or missing inputs: only material issues.
5. Why this structure: 2–4 short bullets.

For Chinese input, give polished English first, then brief Chinese notes on the major structural choices (see references/chinese-author-workflow.md).