SimpleEnglish

Why it's ranked here

The case rests on measurement that is unusually honest. The author publishes a script that recomputes every headline number from committed raw files, and CI runs it on every push. On 16 replies from one Claude model, visible formatting defects fell from 218 to 11. More telling, the author committed an audit admitting the earlier 81% linter win measured rule obedience, not readability, and that an eight-line prompt beat the old skill on what readers see. The skill was rebuilt around that finding. MIT licensed, with a changelog that states the measured effect, or its absence, for each release.

What's good

The rules are specs, not adjectives: a word ceiling per sentence, a banned-modal list, a fixed verb for checking. That makes them testable, and the bundled linter does test them. The file check is advisory and never blocks, and it skips the agent's own memory and configuration files, including through symlinks, so it does not burn tokens on notes nobody reads. The repository ships no text from the standard itself and explains that the dictionary must be built locally from your own copy of the free official PDF. The before and after examples cover READMEs, runbooks, error messages, incident reports and release notes.

Tradeoffs

The published reply figures come from release 2.0.1. The changelog says no benchmark has run on 2.1.0 or later. All runs use one Claude model at low effort with Claude judges, and the authors flag family bias themselves. The short 60-token prompt variant still says five sentences maximum, a cap the 2.1.0 changelog says was removed everywhere. The bundled linter is a regex pass with no passive-voice or part-of-speech detection. The full plugin needs both Node.js and Python 3, and Codex runs only the session-start hook, not the writing checks. The rules strip persuasion, so the skill itself says not to use it on marketing copy.

How to use it well

Install it as the Claude Code plugin if you want the automatic checks. The plain skills installer adds only the rules, and the output style setting does nothing without the plugin. Use it for documentation, runbooks, error text and release notes, where a misread instruction costs something. Ask for strict mode only when you need the controlled vocabulary, and build the word lists from your own copy of the standard. If the file check is noisy on notes or changelogs, exclude those paths with the documented environment variable. It does not certify compliance with the standard, and it is the wrong tool for blog posts or brand copy.

Observed

License

MIT (LICENSE file and plugin manifest)

Packaging

Agent skill folder plus Claude Code and Codex plugin manifests

Install surface

npx skills add, claude plugin install, codex plugin add, or paste-in system prompt

Hook runtimes

Node.js for the session-start hook, Python 3 for the lint hooks (.claude-plugin/plugin.json)

Codex hook coverage

SessionStart only; PostToolUse and Stop checks are Claude Code only (src/hooks/README.md)

Linter method

Regex pass with no passive-voice or part-of-speech detection (evals/ste_lint.py)

Hook behavior

Advisory: Stop always exits 0, PostToolUse exits 2 after the write (src/hooks/lint_hook.py)

Benchmark scope

Published reply figures cover release 2.0.1 on one Claude model at low effort (CHANGELOG.md)

Standard content

No STE dictionary text shipped; word lists built locally from the user's own copy (README.md)