Kiterlin/anti-defensive-writing

Anti‑Defensive Writing – What It Is

Anti‑Defensive Writing is a skill specification for LLM‑powered agents (e.g., OpenAI Codex, Anthropic Claude Code). It teaches the model how to rewrite academic or technical prose so that it removes unnecessary apologies, redundant disclaimers, and stacked hedges while keeping the evidence, methodological limits, and ethical/legal boundaries intact. The result is a more assertive, claim‑forward manuscript that still respects scientific rigor.

How It Works

Skill file (SKILL.md) – The core of the project is a single markdown file that encodes the rewrite rules (diagnosis, pruning, positive projection, uncertainty calibration, structural realignment). Agents load this file as a skill and expose a command‑line shortcut ($anti-defensive-writing …).
5‑stage pipeline – When invoked, the agent runs the text through:
- Functional diagnosis of defensive units
- Removal of uninformative disclaimers
- Conversion of negative statements into positive scope definitions
- Collapse of modal hedge stacks into evidence‑based uncertainty
- Realignment of paragraph structure so the main claim leads.
Zero‑runtime dependencies – Installation is just copying the SKILL.md file into the agent’s skill directory; no Python, Node, or compiled binaries are required.

Getting Started

# Example for the recommended Codex skill directory
curl -fsSL https://raw.githubusercontent.com/Kiterlin/anti-defensive-writing/main/install.sh \
  | sh -s -- --dest "$HOME/.agents/skills"

The same script works for legacy Codex (~/.codex/skills) or Claude Code (~/.claude/skills). After installation you can call the skill directly:

$anti-defensive-writing Revise these paragraphs. Remove unnecessary defensive language and preserve the evidence and necessary qualifications.

Who Might Use It?

Researchers polishing drafts for conferences or journals.
Grant writers who need to present aims with confidence.
Technical writers producing documentation that should sound authoritative.
LLM‑based assistants that help users rewrite text on the fly.

Limitations

The skill does not generate new data; it only rewrites existing content.
It enforces guardrails against hallucination, causal over‑reach, and removal of essential limitations, but the quality of the output still depends on the underlying LLM.
It is a specification – the actual transformation quality varies with the host agent (Codex, Claude Code, etc.).

Quick Reference

Component	What It Provides
`SKILL.md`	Rulebook that the agent follows
`install.sh` / `install.ps1`	One‑line installers for macOS/Linux/WSL and Windows PowerShell
`examples/`	Before‑and‑after rewrites for papers, methods sections, grant proposals
`skill.json`	Manifest used by agents to register the skill

Citation

If you incorporate the skill into your workflow, cite it as:

@software{antidefensivewriting2026,
  author       = {Sunlight and Contributors},
  title        = {Anti-Defensive Writing: Calibrated Assertiveness & Positive Scope for Academic and Technical Prose},
  year         = {2026},
  publisher    = {GitHub},
  journal      = {GitHub repository},
  howpublished = {\url{https://github.com/Kiterlin/anti-defensive-writing}}
}

In a nutshell: this repo supplies a lightweight, agent‑agnostic skill that lets LLMs rewrite scholarly text to be more direct and less defensive, without adding any runtime code beyond a single markdown rule file.