Write Reader-First Papers

中文说明

An instruction-only Agent Skill for revising academic writing so first-time
readers can understand the problem, contribution, evidence, value, and
limitations without decoding internal jargon or formulaic prose.

The skill works with both
Codex and
Claude Code. It follows the open
Agent Skills format and requires no scripts, network services, or
model-specific API.

What it does

• Protects numbers, denominators, citations, terminology, and claim boundaries.
• Separates intentional research and engineering contributions from implementation
  mistakes, debugging effort, and corrective maintenance.
• Reorganizes arguments around the order in which readers need information.
• Rewrites awkward, overly defensive, promotional, or template-like prose.
• Replaces unexplained internal labels with reader-facing language.
• Connects results to their scientific meaning without overstating the evidence.
• Keeps English and Chinese versions aligned in meaning and claim strength.
• Reviews abstracts, introductions, methods, results, conclusions, highlights,
  captions, supplementary information, cover letters, and rebuttals.

This is a writing and reasoning workflow. It does not attempt to evade AI
detectors, disguise authorship, or invent a human writing history.

How it works

The skill applies a reader-first sequence:

1. Establish the source of truth and protect factual content.
2. Map the need, obstacle, proposed work, evidence, value, and boundary, then
   classify claimed advances by their source.
3. Diagnose reader friction through logic, meaning, syntax, evidence,
   contribution attribution, and voice.
4. Revise structure, argument flow, contribution attribution, transitions,
   terminology, sentences, and bilingual alignment in separate passes.
5. Report content, evidence, compilation, rendering, and submission checks as
   separate verification gates.

Detailed rules are loaded progressively from the references/ directory.

Installation

Codex and Codex Cloud

For personal use, copy skills/write-reader-first-papers/ to:

~/.agents/skills/write-reader-first-papers/

For a repository-scoped skill, copy or vendor
skills/write-reader-first-papers/ into the target repository:

.agents/skills/write-reader-first-papers/

Commit the repository-scoped folder so Codex Cloud receives it with the source
checkout.

Claude Code

For personal use, copy skills/write-reader-first-papers/ to:

~/.claude/skills/write-reader-first-papers/

For project use, copy skills/write-reader-first-papers/ to:

.claude/skills/write-reader-first-papers/

The optional agents/openai.yaml file supplies Codex interface metadata.
Claude Code can ignore it; the shared behavior is defined by SKILL.md and
references/.

Usage

Invoke the skill explicitly:

# Codex
$write-reader-first-papers Review the manuscript for argument flow and language.

# Claude Code
/write-reader-first-papers Review the manuscript for argument flow and language.

Both agents may also load it automatically when a request matches the
description in SKILL.md.

Example requests:

• “Rewrite this abstract so a first-time reader can understand the contribution.”
• “Review the full manuscript for unclear transitions and formulaic prose.”
• “Shorten every figure and table caption without losing necessary definitions.”
• “Synchronize the English and Chinese versions without strengthening claims.”
• “Check whether the conclusion explains the value of the work rather than
  repeating numbers.”
• “Check whether the contribution list mistakes our own debugging and bug fixes
  for innovations in the final method.”

Repository structure

.
├── README.md
├── README.zh-CN.md
├── LICENSE
└── skills/
    └── write-reader-first-papers/
        ├── SKILL.md
        ├── agents/
        │   └── openai.yaml
        └── references/
            ├── review-checklist.md
            ├── section-playbook.md
            └── style-rules.md

Contributing

Issues and pull requests are welcome. Keep additions:

• general across research fields;
• independent of a particular manuscript, institution, or model provider;
• concise enough to justify their prompt-context cost;
• focused on reader comprehension and evidence fidelity.

License

MIT