fse-writing-style

FSE Writing Style

Use this when revising the main paper. FSE papers are PACMSE journal articles read by SE

empiricists, so they need a software-engineering contribution stated in the first page and

evidence a reviewer trusts. The failure this skill prevents is a technically fine paper that reads

like a systems demo or an ML result with an SE title glued on.

Revision rules

• Lead with the SE contribution: the problem a practitioner recognizes, why the current state

is inadequate, the contribution (technique and/or finding), the evidence, and what changes for

software engineering.

• State research questions as contracts. Each RQ should name what is measured and how it will

be judged; every RQ must be answered explicitly in the results, and no result should exist

without an RQ it serves.

• Pair every claim with proportional evidence — real subjects, a fair baseline, a statistic

with an effect size, or a qualitative code with agreement — not adjectives.

• Argue threats to validity; do not recite them. Name the construct, internal, external, and

conclusion threats that actually bite this study, and say what you did to bound each. A

boilerplate threats paragraph is a tell that the author has not stressed their own claims.

• Respect the page budget as a design constraint, not a formatting afterthought — the ACM

template is generous but finite, and a study that only fits by shrinking threats or method is

over-scoped.

• Maintain heavy double-anonymity in self-citations, tool names, acknowledgements, funding,

and the data-availability wording.

Empirical-SE paper skeleton

| Section | Job it must do | Common failure |

|---|---|---|

| Intro | Problem, inadequacy, contribution, evidence preview, SE payoff — first page | Leads with a technology trend, not a problem |

| Background/Motivation | Why a practitioner or the field needs this now | Motivation by assertion, no grounding |

| Approach / Study design | The technique or the RQs + protocol, reproducibly | Method described too thinly to re-run |

| Evaluation | Each RQ answered with proportional evidence | Metrics that proxy for the claim rather than test it |

| Threats to validity | The threats that bite, each bounded | Generic list untethered from this study |

| Related work | Delta-first positioning against SE literature | Catalog of citations with no contrast |

Sentence-level rewrites

| Draft pattern | FSE-safe rewrite |

|---|---|

| "Our tool significantly improves..." | "detects X% more defects (95% CI ...) than <baseline> on <N real projects>" |

| "We evaluate on a large dataset." | "We evaluate on <N> projects sampled by <criterion>, listed in the artifact" |

| "Results show our approach works well." | "RQ2: developers acted on <fraction> of comments (effect size ...); threats in §5.2" |

| "State-of-the-art performance." | Claim scoped to the subjects, metrics, and regime actually tested |

| "The LLM understands the code." | "the model's output matched <criterion> in <fraction> of cases" |

Threats-to-validity discipline

[Construct]   does the metric measure the SE outcome you claim? (e.g. proxy for causation)
[Internal]    could something other than your technique explain the effect? (confounds, tuning)
[External]    to which projects/languages/teams does the finding generalize?
[Conclusion]  are the statistics appropriate; multiple comparisons corrected; variance reported?
-> for each that bites: state it, then state the mitigation, next to the affected result

Vignette: compressing a three-RQ study

A draft with three RQs, nine figures, and a sprawling background: keep all three RQ answers, the

two figures that carry the headline findings, and a threats subsection per RQ; move secondary plots

and full protocol tables to the artifact with explicit forward references; cut background to what

the argument needs. The test of a good cut: a reviewer should be able to answer "what did each RQ

find, and what threatens it?" from the body alone.

Output format

[Writing diagnosis] clear / under-motivated / over-claimed / evidence-mismatched / over-scoped
[First-page fix] <new framing leading with the SE contribution>
[RQ audit] <RQ -> metric -> where answered -> proportional? yes/no>
[Threats fix] <threat that bites -> mitigation to add, placed by the result>
[Anonymity edits] <tool names / self-citations / links to rewrite>