Anti-Defensive Writing: A Claude Skill for Stronger Academic Paper Sections

Defensive Writing as a Specific, Fixable Problem

Defensive writing in academic papers is a pattern where the author preemptively concedes weaknesses before any reviewer raises them. The README gives concrete examples: writing 'although our method improved accuracy on two in-domain datasets, we must acknowledge we have not tested cross-domain generalization, so these results should be interpreted with caution' instead of simply stating 'our method improved accuracy on two in-domain datasets; cross-domain generalization has not been evaluated.'

This pattern appears in abstracts, introductions, experiment sections, and conclusions. In each location it performs a similar function: it signals to the reader that the author has identified a gap and is managing the reviewer's reaction in advance. The README frames this as handing the reviewer a knife alongside a user manual.

The project is aimed at academic researchers, especially those whose papers are written in Chinese and then submitted to international venues, though the repository provides both a Chinese skill and an English one. It is not a grammar or style corrector. The problem it addresses is argument structure and the selection of what to foreground.

The Press Conference Rule and the Twelve Supporting Rules

The central principle in the repository is what the README calls the press conference rule: a paper is a launch event, not a project summary. A launch event covers one thing, the strongest advantage. Metrics the method cannot win are not framed as competitions. Results where the method does not lead are not described as losses. Each experiment carries a specific argumentative function rather than serving as an exhaustive results archive.

The twelve rules are grouped into four areas. Narrative rules cover organizing around advantages rather than writing work-report-style summaries, not setting up unfavorable comparisons, stating advantages explicitly, controlling the scope of comparisons, and allowing full restructuring of the story. Language rules cover eliminating self-weakening phrases and not framing local phenomena as indictments of the whole method. Experiment rules require that each experiment carry a specific argumentative role. Structure rules address the abstract and introduction as the opening of the launch event, and the conclusion as a memory reinforcement section rather than a self-cancellation.

The README provides a comparison table showing what each rule changes. A conclusion that suddenly reverses the paper's argument in the final paragraph is one of the four structural patterns the rules specifically target.

Installing the Skill and Using the Standalone Prompt

The repository offers two installation paths. The first is a skill for Claude Code, Claude Cowork, Cursor, or other agent tools that support skill directories. Copy the appropriate language directory into your Claude skills folder:

# 中文版
cp -r skills/anti-defensive-writing ~/.claude/skills/

# English version
cp -r skills/anti-defensive-writing-en ~/.claude/skills/

After that, ask Claude to edit a paragraph or abstract and it applies the twelve rules automatically without requiring any additional instructions in the chat.

The second path requires no installation at all. The repository includes a standalone prompt file at prompts/quick-prompt-en.txt. Copy its contents into any chat with any AI assistant and paste your paper section after it. The README describes this as a zero-friction path for people who do not use Claude Code or similar tools. The Chinese version is at prompts/精简版提示词.txt.

Both paths apply the same underlying rule set. The skill version has the advantage of triggering automatically in future sessions without needing to re-paste instructions.

Repository Layout and What Each File Contains

The repository has a flat, purposeful structure. The top level contains README.md and README_EN.md, a prompts/ directory with the two standalone prompt files, and a skills/ directory with the two skill subdirectories.

The skills/anti-defensive-writing/ directory holds SKILL.md, the Chinese version of the full rule set and decision flowchart. The skills/anti-defensive-writing-en/ directory holds the equivalent English version. The README notes that the complete twelve rules and decision process are in SKILL.md, while the README itself provides only a summary.

The prompts/ directory contains exactly two files: the condensed Chinese prompt for immediate use in any AI chat, and quick-prompt-en.txt for the equivalent English use. These are intended for cases where someone needs to use the principle once without committing to installing the skill.

There are no configuration files, no package manager files, and no build steps. The repository is text-only and has zero runtime dependencies, which is consistent with its purpose as a set of instructions rather than a software library.

The Four Section Types It Is Designed to Fix

The README identifies four specific locations in a paper where defensive writing concentrates. The first is the abstract and introduction, where work-report framing positions the paper as a sequence of things that were tried rather than as a claim supported by evidence.

The second is the experiment section, which in defensive writing becomes an appendix-style dump of all results without each experiment carrying a specific argumentative role. The README frames each experiment as a tool for proving one thing, not a display of thoroughness.

The third is the conclusion. Defensive writing conclusions often introduce new limitations or self-criticisms in the final paragraph, which the README characterizes as a self-cancellation that undermines the paper's entire argument. The rule for conclusions is to reinforce only the key memory points: what was solved, what was proved, and why it matters.

The fourth is the rebuttal. The README mentions using the press conference principle as a self-check before writing a rebuttal, specifically to identify which admissions would hand the reviewer new objections.

The skill is not designed to help with tasks that are not argument structuring, such as summarizing an article, writing a tweet, or composing a literature review that must represent the field fairly.

What the Skill Does Not Address and Where It Falls Short

The twelve rules operate entirely on argument structure, rhetorical framing, and word-level hedging patterns. They do not address citation formatting, grammar, sentence clarity, vocabulary selection for a specific journal style, or the empirical validity of the claims being made. A paper that makes incorrect claims will not become correct by removing defensive hedges.

The skill also does not enforce factual accuracy. If a researcher applies the press conference rule to a set of results that actually do show a limitation in all tested conditions, the skill will frame those results more confidently, but the underlying experimental gap remains. The README explicitly acknowledges that setting up unfavorable comparisons as non-competitions removes them from the argument, not from the data.

A plain system prompt telling an LLM to 'write like a press conference and cut all hedging' is a functional alternative for a one-off session. The advantage of this repository over a plain prompt is the twelve explicit, named rules and the decision flowchart in SKILL.md, which give the model a repeatable, inspectable framework rather than a single vague instruction. For users who need to apply this consistently across many paper sections over time, the skill installation path provides that consistency.

License and Maintenance

The repository is MIT licensed. The README invites pull requests and sharing among research group members. There are no GitHub releases and no version numbering in the repository, reflecting its nature as a text-based instruction set rather than versioned software.

The last push was on 2026-09-12. The two skill files in SKILL.md format represent the entire deliverable of this repository. Updates to the rules or prompts would appear as changes to those files. The README's reference to the author contact on GitHub and the invitation for PRs suggests the rules are considered a live document rather than a finished specification.

Editorial conclusion

Researchers who recognize their own paper sections in the defensive column of the comparison table in the README are the direct audience. Installing the English skill under ~/.claude/skills/ requires only a single cp command and then triggers automatically when asking Claude to edit abstract or introduction text. The repository does not address citation formatting, bibliographic style, or grammar correction; it targets only the hedging patterns that weaken an argument. The last push was on 2026-09-12.

Frequently asked questions

Official sources

1. Adkid-Zephyr/anti-defensive-writing-Skill on GitHub
2. Issues
3. License: MIT
4. README