anti-defensive-writing-Skill: A Prompt Pack That Stops You Apologising in Your Own Paper

The problem it names: defensive writing

The README opens with a scene most graduate students recognise. You run a paper through an AI, and the result reads like a confession: hedges in every clause, a limitations paragraph that arrives before the reviewer asks for one, and a structure that walks through what you tried in chronological order. The repository calls this defensive writing and argues it hands the reviewer the knife along with instructions for use.

The target user is narrow and specific. It is someone who has results already and is writing or revising the abstract, introduction, conclusion, experiment section, or a rebuttal. The README lists exactly those four scenarios. It is not aimed at people who need help finding a research question, and it is not a reference manager or a LaTeX template. The whole repository is a skill directory plus two prompt files, which is why the README calls it lightweight and zero-dependency.

One principle, twelve rules, and a decision flow

The mechanism is not a model. It is a written rule set that an AI agent loads as context. The README states the core principle as a single sentence: a paper is an academic launch event, not a project summary, and a launch event talks about one thing, your strongest advantage.

The twelve rules are grouped into four buckets. Narrative rules say to organise around the advantage, avoid work-report structure, refuse to make a losing metric into a competition, state the advantage explicitly, control the scope of comparison, and allow a full rewrite of the story. Language rules ban self-undermining expressions, forbid calling a non-winning result a loss, and forbid generalising a local phenomenon into a rejection of the whole method. One rule covers experiments: every experiment must carry an argumentative duty. The structure rule says the abstract and introduction are the launch-event opening and the conclusion only reinforces the memory point.

The README's comparison table shows the intended edit. A defensive sentence admits that cross-domain generalisation was not tested and asks the reader to interpret results cautiously. The launch-event version states the accuracy gain, then notes in a separate clause that cross-domain generalisation has not been evaluated. Same fact, no apology attached. The full rule set and the decision flow live in skills/anti-defensive-writing/SKILL.md, which the README points to rather than reproducing.

Installing the skill under ~/.claude/skills/

There are two ways in, and the README presents the prompt route first because it needs no installation at all. Open prompts/精简版提示词.txt for Chinese or prompts/quick-prompt-en.txt for English, copy the contents, and paste them at the top of your conversation with any AI tool. Then send the paragraph you want rewritten. The README describes this as copy-and-use with no threshold.

The second route is for agent tools that read a skills directory, which the README names as Claude Code, Cursor and other agent tools. Copy the language directory you want into your 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, the README says you can ask the AI to revise a paper paragraph or polish an abstract and it will work according to the launch-event principle. What you should see is a rewritten passage with hedges removed and the claim stated up front, not a list of suggested edits. If your tool does not read a skills directory, the prompt route is the only one that applies, and the README does not describe a plugin format or a package manager install.

What it will not do, and when it is the wrong tool

The rule set is deliberately opinionated, and that is also its failure mode. Telling a model to refuse to make a losing metric into a competition is useful when you have a genuine advantage and dangerous when you do not. If your method is slower than the baseline on the metric your venue cares about, the skill's rules push the text away from saying so plainly. The README's own examples keep the unfavourable fact in the sentence; they only strip the apology. A user who reads the rules as permission to omit the fact has misread them, and nothing in the repository enforces that distinction.

Second, this is a rewriting tool, not a discovery tool. It cannot tell you which of your results is the strongest advantage, and the README's first narrative rule assumes you know. Third, the repository does not document rollback, version pinning or any evaluation of whether the skill improves acceptance. There are no releases retrieved, so there is no changelog to diff between versions. If you need a tool that checks citations, enforces a journal style file, or detects plagiarism, this is not it.

How it differs from a style checker or a general writing prompt

A style checker such as a journal's language-editing service or a grammar tool works at the sentence level: passive voice, tense, article usage. The anti-defensive-writing skill works at the argument level. Its rules are about which comparison you set up, which experiment earns a place in the paper, and whether the conclusion reinforces a memory point or quietly retracts one. A grammar tool would leave the defensive sentence in the README's table untouched, because it is grammatically fine.

The closer alternative is a generic "improve my academic writing" prompt. The difference is that this repository fixes its stance in writing. A generic prompt lets the model decide what good writing means on each run, which is how you get hedges back. Here the twelve rules are stored in SKILL.md and travel with the skill, so the same stance applies across sessions. The cost of that fixed stance is the rigidity described above: you get one editorial philosophy, not a menu.

Licence and the real cost of adopting it

The repository is MIT licensed, which permits commercial use, modification and redistribution provided the copyright notice and permission notice are kept. The README says PRs are welcome. Because the artefact is a markdown skill file and two text prompts, there is no dependency graph to audit and no runtime to patch. Upgrading means replacing the copied directory or re-copying the prompt text, and the README does not describe a versioned update path, so you are responsible for noticing changes yourself.

The maintenance signal is the last push on 2026-09-12, four days before this writing, and the repository is not archived. That is recent activity, but there are no releases, so there is no tagged version to pin. If you fork it to match your field's conventions, that fork becomes the thing you maintain, and the MIT terms let you do that without asking.

Editorial conclusion

Adopt it if you already have results and keep undermining them in the abstract, introduction or conclusion, and if you write in Chinese or English. Do not adopt it if your problem is missing evidence, a failed experiment you have not decided how to report, or a thesis chapter that needs a methods description rather than an argument. Before relying on it, open skills/anti-defensive-writing/SKILL.md and check whether its twelve rules match your venue's conventions on limitations sections; some reviewers expect an explicit threats-to-validity paragraph, and the launch-event framing pushes against that. The prompt files under prompts/ are the cheapest way to test the rules on one paragraph before you copy anything into ~/.claude/skills/.

Frequently asked questions

Official sources

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