# Scientific Prose

> English scientific report writing and revision pass. Use for research reports, progress reports, figure-heavy PDFs, manuscripts, rebuttals, technical summaries, and slide text that must keep evidence, uncertainty, captions, and conclusions scientifically defensible without AI-sounding or defensive/self-undermining prose.

- Skill: `yuukias/scientific-prose` (Agent Skill, multi-file: 5 files)
- Install (CLI): `npx skillmds@latest add yuukias/scientific-prose`
- Raw SKILL.md: https://api.skillmd.com/api/skills/yuukias/scientific-prose/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Docs & Writing
- License: MIT-compatible synthesis from cited public/academic writing guid
- Author: YuukiAS (https://skillmd.com/u/yuukias)
- Updated: 2026-09-21
- Page: https://skillmd.com/skills/yuukias/scientific-prose

---

# Scientific Evidence Prose

Use this skill as the final English writing pass for scientific reports and research communication. It is not a layout tool. It makes prose scientifically defensible, readable, and less template-like.

This is not the manuscript planner, citation verifier, Chinese final pass, or an
AI-detector evasion tool. Use it after the evidence and intended section are
known, and keep `writing-fidelity` active when exact numbers, labels, equations,
citations, source wording, or author-approved structure must be preserved.

## Use When

- Writing or revising English research reports, progress PDFs, lab updates, manuscripts, thesis sections, figure captions, or technical summaries.
- Turning experiment notes, ablations, error analysis, tables, or figures into report prose.
- Checking whether a result statement overclaims beyond the evidence.
- Making slide/report text sound like a researcher describing evidence, not a generic AI summary.
- Removing defensive or self-undermining scientific phrasing while preserving exact limitations.

Do not use it for pure citation search, document conversion, journal formatting, whole-paper structure planning, reviewer-style risk assessment, or Chinese "say it plainly" final passes. Use `research-paper-workflow` for paper-level structure, `citation-verification` for source/claim support checks, `chinese-prose` for Chinese reader-facing prose, and the relevant PDF, DOCX, PPTX, or venue skill for file or formatting work.

## Core Rule

Every paragraph should answer one reader question:

1. What question is being addressed?
2. What evidence is available?
3. What can be concluded from that evidence?
4. What remains uncertain?
5. What decision or next step follows?

If any part is missing, keep the wording bounded rather than filling the gap with confident prose.

## Workflow

1. Identify the document type: report, progress update, manuscript section, figure caption, slide text, or executive summary.
2. Protect the evidence ledger: numbers, metrics, cohorts, baselines, sample sizes, methods, dates, versions, figure labels, statistical qualifiers, and source attributions.
3. Mark claims by strength:
   - `observed`: directly shown in the supplied result.
   - `supported`: follows from multiple supplied observations.
   - `hypothesis`: plausible but not yet tested.
   - `speculation`: useful idea, but not evidence-backed.
4. Rewrite so confidence matches claim strength.
5. Remove AI-like prose patterns: grand openings, vague importance claims, "this highlights", "underscores", unsupported "significant", generic future-work endings, and symmetrical three-item filler.
6. Check information flow: old/context information first, new/emphasized result last.
7. For figures, check that captions and body text do different jobs.
8. Return the revised text. If the input is risky, add a short "Evidence notes" section listing overclaims or missing evidence.

Do not erase real limitations to make the text sound more confident. If the
limitation is real, make it exact and proportional: where it applies, what still
holds, and what evidence would be needed to go further.


## Confident, Non-Defensive Scientific Framing

Use this pass when a draft sounds apologetic, self-undermining, or organized like a lab diary rather than an evidence-backed scientific argument. The goal is confident precision, not overclaiming.

1. Identify the strongest claim that the supplied evidence can honestly support.
2. Organize the paragraph around problem, gap, approach, evidence, and significance rather than chronology of attempts or failures.
3. Classify material as central proof, supporting context, required limitation, optional detail, process history, failed attempt, or distraction.
4. Remove unnecessary self-weakening language such as "unfortunately", "only", "merely", "fails to", "not ideal", or broad labels like "severe weakness" unless the evidence truly requires that judgment.
5. If a result is mixed, narrow the claim, change the evaluation dimension, or present the result as a scoped tradeoff. Do not set up a contest the work does not need to win.
6. Keep required limitations exact and proportional: what is limited, where the evidence still holds, and what would be needed to extend the claim.
7. In reviewer responses, admit real errors cleanly, describe the fix, and return to evidence. Avoid over-apology and avoid conceding claims the paper never made.
8. Before returning a revision, check whether any sentence gives readers a broader attack surface than the evidence requires.

Source distilled from `Tina0514/prevent-defensive-ai-writing-skill` at `9686ccbacad5`; see `docs/provenance/WRITING_SKILL_REPOS_2026_07_28.md`.

## Report Paragraph Pattern

Prefer this order in English scientific reports:

```text
[Question/context]. [Specific evidence and comparison]. [Interpretation with boundary]. [Uncertainty or competing explanation]. [Decision or next step].
```

Keep it in normal prose. Do not mechanically label every sentence unless the user asked for a checklist.

Avoid formulaic three-part summaries and generic transitions when the evidence
does not need them. The structure should follow the scientific argument, not a
fixed "first/second/finally" template.

## Claim Boundaries

Use bounded verbs when evidence is incomplete:

- "suggests", "is consistent with", "we observed", "in this subset", "under this setting"

Avoid stronger verbs unless evidence supports them:

- "proves", "demonstrates" for single runs or unreplicated experiments.
- "significant" without a statistical or practical definition.
- "robust" without sensitivity, subgroup, or external validation.
- "generalizes" without external data or held-out conditions.

## Figure Captions

For figure-heavy reports, captions should usually contain:

1. What is being compared.
2. Under which condition, cohort, model, or dataset.
3. Which metric or visual encoding matters.
4. The main observation.
5. Abbreviations, panels, error bars, and sample sizes if needed.

Do not dump the full method into the caption. Put method differences in nearby prose if they explain the result.

## English Style Pass

Use direct, concrete sentences:

- Replace vague openings with the actual research question.
- Replace "This highlights the importance of..." with the observed consequence.
- Replace "Future work should explore..." with the next test and the decision it will enable.
- Prefer active voice when responsibility matters: "We reran the ablation" instead of "The ablation was rerun".
- Keep technical terms if they are the right terms. Removing expertise is not the goal.

## References

Read `references/scientific-report-checklist.md` when doing a full report pass. Read `references/source-notes.md` when provenance matters.