skills

This repository contains Deliberate Writing, a Codex skill for drafting and revision based on communicative purpose. It asks the model to settle meaning and discourse strategy before wording, then permits revisions only when they fix a concrete defect.

The skill is language- and genre-agnostic. It does not use phrase blacklists, generic style targets, or language-specific patches.

Install

Clone the repository:

git clone https://github.com/UlloaSP/skills "C:\path\to\skills"

Link the skill into the Codex skills directory:

New-Item -ItemType Junction `
  -Path "$env:USERPROFILE\.codex\skills\deliberate-writing" `
  -Target "C:\path\to\skills\skills\deliberate-writing"

Restart Codex after installation. Invoke it explicitly with $deliberate-writing, or let Codex select it automatically.

Evaluate

The experiment compares a control response with a response produced under the skill instructions:

.\scripts\run-experiments.ps1

The script writes raw responses and a blind rubric-based assessment under experiments/results/. Model output varies, so the checked-in report records the exact run rather than claiming a permanent score.

The default evaluation sandbox is read-only. If your Codex configuration requires loading a global mandatory skill and blocks that read, review the benchmark inputs first and run with -Sandbox danger-full-access.

Validate the package structure with:

uv run --with pyyaml python .\scripts\validate_skill.py

Repository topics

codex-skill, writing, ai-writing, discourse, rhetoric, natural-language-generation, prompt-engineering, llm

Status

This is an experimental v0.1. Its central claim is narrow and falsifiable: planning for purpose, diagnosing before revision, and stopping when no functional defect remains should outperform generic drafting or open-ended "improvement" prompts across different kinds of prose.