SimpleEnglish Agent Skill Enables LLMs to Write in ASD‑STE100 Simplified Technical English
TL;DR – What the SimpleEnglish skill does and why it matters
The SimpleEnglish agent skill forces large language models to generate text that follows the ASD‑STE100 Simplified Technical English standard, eliminating most "AI slop" and reducing sentence length, token count, and rule violations by roughly 73% across six Claude models. This makes AI‑generated documentation clearer, less ambiguous, and safer for mission‑critical environments such as aerospace maintenance.
Core functionality – an agent‑level plug‑in, not a prompt hack
The project provides a single‑file Agent Skill that can be installed in any environment supporting the Agent Skills standard (Claude Code, Cursor, VS Code Copilot, OpenAI Codex, Gemini CLI, Goose, OpenCode, etc.). Installation is a one‑liner:
npx skills add AminBlg/SimpleEnglish
The skill injects a system prompt derived from the 53 numbered rules of ASD‑STE100, ensuring that every generation is constrained to:
- ≤ 20 words per instruction, ≤ 25 per description
- One meaning per word (no ambiguous synonyms)
- Simple present/past tenses only
- No "‑ing" verb forms
- Active voice only
- No hedging modal verbs except can, will, must
- Condition before command, one instruction per sentence, etc.
Because the skill lives in the agent runtime, it works consistently across tools without requiring users to remember a custom prompt.
Measurable impact – benchmark results
The repository includes a reproducible benchmark (96 runs: 6 Claude models × 8 writing tasks × 2 conditions). Key findings:
| Model | STE violations / 100 words (baseline) | STE violations / 100 words (skill) | Reduction |
|---|---|---|---|
| claude‑opus‑4‑8 | 1.05 | 0.62 | 41 % |
| claude‑opus‑4‑7 | 2.28 | 0.42 | 82 % |
| claude‑opus‑4‑6 | 2.24 | 0.40 | 82 % |
| claude‑opus‑4‑5 | 2.55 | 0.57 | 78 % |
| claude‑sonnet‑5 | 2.67 | 0.53 | 80 % |
| claude‑sonnet‑4‑6 | 2.06 | 0.52 | 75 % |
All models also produced fewer output tokens and shorter average sentence length (11.2 → 9.7 words). The evaluation used a deterministic regex linter that applies the same rule set to both baseline and skill‑augmented outputs, guaranteeing a fair comparison.
Real‑world examples – before and after
The README showcases side‑by‑side rewrites. A typical LLM output:
"Leveraging sqlpipe's robust architecture, users can seamlessly synchronize their Postgres tables to S3 with minimal configuration overhead. ..."
After applying the skill:
"sqlpipe copies your Postgres tables to S3. It needs one configuration file. Before you start, make sure that your AWS credentials are correct. If they are not, S3 rejects the upload with a permission error."
Other examples include error messages, incident reports, and release notes, all transformed into concise, imperative prose that a tired mechanic could follow without ambiguity.
Installation pathways – CLI and web UI
- CLI –
npx skills addworks for any Agent Skills‑compatible harness. - Claude.ai – Upload the
SKILL.mdfile via Settings → Customize → Skills. - ChatGPT – Paste the system‑prompt block into Custom Instructions.
- Gemini – Add the same block to a Gemini "Gem"'s instructions.
- Any other chatbot – Paste the prompt and request the skill be applied to all subsequent generations.
A minimal ~60‑token version is also provided for contexts where token budget is tight.
Community feedback – common themes from Hacker News
"How does it affect the model's intelligence?" – A user wondered whether constraining output might limit the model's reasoning. The skill only modifies the surface language; the underlying model still performs the same reasoning steps.
"One line in the system prompt is enough" – Several commenters argued that a simple prompt (e.g., "Use ASD‑STE100") can achieve comparable results. The author counters that the skill encodes the full 53‑rule set, guaranteeing consistent enforcement across agents without manual prompt engineering.
"I got good results by just prefixing the request" – A user reported acceptable rewrites by manually asking for STE, but noted that the skill eliminates the need for ad‑hoc prompting and provides deterministic rule compliance.
"It’s cruft; capable users don’t need it" – Some dismissed the skill as unnecessary boilerplate. Others highlighted its value for non‑technical writers or for enforcing style in large codebases where manual prompting is impractical.
"The README itself violates STE" – A tongue‑in‑cheek observation that the project's own documentation breaks several STE rules, underscoring the difficulty of writing in the style.
"Use linters for post‑processing" – A suggestion to treat STE enforcement as a linter rather than a generation constraint. The SimpleEnglish skill can be seen as a pre‑generation linter, ensuring the model never emits prohibited constructions.
Overall, the discussion reflects a split between purists who favor minimal prompting and practitioners who appreciate a reusable, standards‑compliant enforcement layer.
Limitations and scope
- Not for marketing or brand voice – The skill deliberately blocks promotional language; it is intended for technical documentation, error messages, runbooks, incident reports, and release notes.
- No official certification – ASD‑STE100 does not certify tools; the skill provides pragmatic compliance but does not claim formal certification.
- Model‑agnostic but evaluated only on Claude – While the skill works with any Agent Skills‑compatible model, benchmark data currently exist only for Claude families.
- Potential drift – As models evolve, the skill may need updates to handle new tokenization quirks or emergent phrasing patterns.
How to extend or adapt the skill
- Domain vocabularies – Users can augment the skill with a custom glossary to retain industry‑specific terminology while still obeying STE syntax.
- Strict mode – The repository includes a stricter configuration that more closely mirrors the official rule set, useful for environments that require near‑certified output.
- Integration with linters – Pairing the skill with a post‑generation English linter (e.g., Vale) can catch any residual violations.
License and provenance
The code is released under the MIT license and contains only paraphrased rule descriptions; no proprietary ASD‑STE100 text is reproduced. The project is unaffiliated with the ASD organization.
Bottom line
The SimpleEnglish agent skill demonstrates that a lightweight, standards‑based plug‑in can dramatically improve the clarity and safety of AI‑generated technical documentation. By embedding the 40‑year‑old ASD‑STE100 rule set directly into the generation pipeline, it removes the need for ad‑hoc prompting, reduces token usage, and yields output that is consistently concise and unambiguous—qualities essential for high‑stakes domains like aerospace maintenance.