Show HN: STE80 Humanizer – de-AI text by writing 80% of the way to ASD-STE100
TL;DR Highlight
A skill strips 26 AI tells and rewrites text in short ASD-STE100-style sentences
Who Should Read
Developers whose LLM-written docs, READMEs or runbooks are long, wordy and obviously AI-sounding
Core Mechanics
- Two passes, not one
- Pass 1 removes the 26 AI tells from Wikipedia's 'Signs of AI writing'.
- Pass 2 rebuilds sentences 80% of the way to ASD-STE100.
- Removing phrases alone leaves long sentences that carry three ideas, per the README.
- ASD-STE100 is controlled aircraft-maintenance English
- It came from the 1980s aerospace industry for mechanics whose first language is not English.
- Rules: 20 words per instruction, 25 per description, one instruction per sentence, active voice, simple tenses.
- It has about 900 approved words and at most six sentences per paragraph.
- 80% keeps structure and loosens vocabulary
- The default 80% skips the approved dictionary and uses plain common words, with check and test allowed as verbs.
- It keeps may and should for real uncertainty and allows contractions in casual text.
- Strict 100% bans all of these.
- Karpathy's tip is the starting point
- The README says Karpathy posted on 2 October 2026 to ask an LLM to explain things in ASD-STE100.
- He sometimes asks for '80% of the way' because the spec is stringent.
- Fidelity is the difference from a plain prompt
- In the authors' small test, the run without the skill invented a quote and changed '40+ customers' to '40-odd'.
- The run with the skill kept every fact and flagged a contradiction in a README source.
- The authors call it a hint, not proof.
- Install and linter are included
- npx skills add Dodaios/ste80-humanizer installs it for Claude Code, Codex, Cursor and others.
- scripts/ste_check.py is a dependency-free Python linter for sentence length, passive voice, semicolons and more.
- It is neither an AI detector nor a compliance checker.
Community Reactions
- The 'humanizer' name misleads, one commenter said
- It removes AI tics and adds no human voice, in their view.
- The goal is clearer, more concise text, not fooling readers.
- A commenter asked how it differs from similar tools
- They had tested SimpleEnglish (AminBlg/SimpleEnglish) across several harnesses and said it works well.
- Even a bare 80% request helped, one user reported
- Asking for 80% of the way to ASD-STE100 made output much more readable.
- They felt less brain fog reading Opus output.
- Alternatives and skepticism appeared too
- One commenter suggested ISO 24495-1 for technical writing.
- Another wondered whether it defeats text watermarking.
- Others said the README itself reads like AI and that people should just learn to write.
How to Apply
- Rewrite one AI-written doc at 80% today
- Pick a recent LLM-written README or internal doc and paste it into the prompt below.
- Compare with the original to check no facts were dropped or changed.
- Install the skill for your doc work
- Run npx skills add Dodaios/ste80-humanizer in your project.
- Ask your agent: 'Rewrite docs/intro.md 80% of the way to ASD-STE100'.
- Run the linter on your docs
- Use python scripts/ste_check.py docs/runbook.md to find long sentences, passive voice and semicolons.
- Treat warnings as hints because it is a heuristic, not a gate.
- Run runbooks and safety steps in Strict mode
- Ask for 'strict STE' on procedures to apply the 100% rules.
- Skip fiction, speeches and brand copy, or use Pass 1 only.
Try It
# Paste into Claude or ChatGPT as is (simplified version of the STE80 Humanizer method)
You are an editor. Fix the {{document}} below in two passes.
Pass 1: Remove AI tells. Cut chatbot openers like 'Great question!', 'not just X, it's Y', one-line closers, dash overuse, stacked hedges like 'could potentially', and inflated significance.
Pass 2: Rewrite 80% of the way to ASD-STE100. One idea per sentence, about 20 words for instructions and 25 for descriptions, active voice, simple tenses, condition first (If X, do Y), procedure apart from description, one topic per paragraph with at most six sentences.
You do not need the approved dictionary. Keep may and should only for real uncertainty.
Rules: Keep every fact and number exactly ('40+' stays '40+'). Add nothing. Flag any contradiction in the source.
Output: the rewritten text, then a 3-line summary of changes.
{{document}}
Next: check that numbers and may/should match the original, then ask for 'strict STE' and compare.Terminology
Related Papers
How Specification Explicitness Shapes Failure in LLM-Generated Code
Spelling out invariants, edge cases and constraints cut hidden-test failures in LLM code by 47.2%.
Same Task, Different Statement: When and How Reformulation Helps LLM Problem Solving
Rewriting problem statements to fit their type raises LLM accuracy, but blind rewriting hurts.
Character-Name Concentration in LLM Fiction: Measuring Displacement and Mitigating Repetition
Drafting name candidates and filtering by frequency cut LLM character-name repetition by 95%
LLM Voice
Martin Fowler: LLM-drafted prose keeps an uncanny tone even after human editing.
The Instruction Tax on Factual Recall: Base Models Can Beat Their Instruction-Tuned Counterparts
Few-shot completion with repeated sampling on a base model beat chat: F1 0.684 vs 0.568
AI is a tool and using it well takes skill
A rebuttal to Hickey: LLMs are tools, and skill in directing and checking them decides the outcome
Reflective Prompt Optimization of a Deep-Research Agent: A Topic-Disjoint, Multi-Seed Protocol for Measuring Transferable Gains and Prompt-Level Overfitting
Reflective prompt optimization raised held-out citation reward from 0.702 to 0.781 on a deep-research agent.
GPT‑6 and Intelligent UI for everyone
GPT-6 reportedly adds generated UI to chat; with no article text, only comment reactions are available.