A new open-source agent skill aims to make AI-generated technical documentation follow principles adapted from ASD-STE100 Simplified Technical English, a controlled language associated with aerospace documentation. The SimpleEnglish project packages its instructions in a dependency-free folder and publishes the material under the MIT licence.
The repository's stated goal is narrower than general writing improvement. It targets documentation, error messages, incident reports, release notes and related technical material where ambiguity can create operational problems. The project explicitly excludes marketing and brand voice from its intended scope. Its examples replace indirect, promotional or vague wording with shorter statements that name an action, condition or failure directly.
SimpleEnglish contains 53 numbered rules in nine sections, according to the repository. The rules are paraphrased rather than reproduced from the official standard, and the author says the project includes no ASD-STE100 dictionary content. The repository describes itself as unofficial and says it is neither affiliated with nor endorsed by ASD or STEMG. It also warns that using the skill does not make output certified.
Installation options vary by agent environment. The project says it works in systems that support the Agent Skills format, including Claude Code, Cursor, VS Code Copilot, OpenAI Codex, Gemini CLI, Goose and OpenCode. For chat systems without direct skill support, it offers longer and compact prompt versions that can be added to persistent instructions. The repository also includes an output-style option for Claude Code.
The author reports testing the skill across seven models and eight writing tasks, producing 112 generations. A deterministic linter measured 74.6% fewer violations per 100 words with the skill enabled, the repository says. In a blinded pairwise assessment covering 56 pairs, the skill-assisted version was preferred 45 times, tied five times and lost six times. The reported mean rubric score rose from 6.0 to 8.1. A separate run across four models also reduced measured violations in every case, according to the published summary.
Those figures are project-reported results, not independent certification. The repository provides benchmark methods, outputs and reproduction commands so users can inspect the evaluation design. That transparency is important because a regex linter and a model-based judge measure particular definitions of clarity rather than every form of real-world comprehension.
The release reflects a broader effort to constrain AI writing with explicit, auditable specifications that teams can inspect, modify and test in their own documentation workflows. Instead of asking a model to be clear, SimpleEnglish gives it measurable limits and structural instructions, while leaving domain vocabulary available when technical accuracy requires it.
