“Created in the late 1970s.”
The project began in 1979. The first AECMA Simplified English Guide was published in 1986. ASD-STE100 is its present name; Issue 9 is dated 15 January 2025.
Source: official STE history ↗THE EVIDENCE LEDGER / CHECKED 19 OCT 2026
A useful idea.
With the edges intact.
Krakem started from an argument for clearer AI output: borrow rules from Simplified Technical English, and prefer visual or interactive explanations to plain text. Quoted headings below are that argument’s claims. Each verdict says what holds up, what needs qualification, and what we could not confirm.
PRIMARY SOURCES WHERE AVAILABLE01SIMPLIFIED TECHNICAL ENGLISH
Precision, without borrowing certainty.
“Under twenty words. One action. One name.”
Rule 5.1 permits a maximum of 20 words in procedural sentences. Rule 5.2 separates instructions per sentence, with an exception for simultaneous actions. Rule 1.11 requires the same technical noun for the same item. “Under 20” and “always one action” are too absolute.
Source: Issue 9, rules 5.1, 5.2 and 1.11 (PDF) ↗A short prompt is not STE certification.
STE includes a controlled dictionary and a complete set of writing rules. Borrowing a few principles can guide an edit; full compliance requires more. The official STE group also warns that AI assistance cannot replace human oversight.
Official FAQ ↗STE and AI white paper (PDF) ↗“AI is bad at technical tasks because the writing is murky.”
Unclear prose is one failure mode. Incorrect facts, missing context, faulty reasoning, and untested code are others. Clearer wording makes an answer easier to inspect; it does not make the answer correct.
Editorial assessment. The suggested directive is a technique to try, not a measured performance guarantee. The aerospace safety claim is not used here as proof that AI synonym changes have caused deaths.
02EPHEMERAL TOOLS & VISUAL EXPLANATIONS
Choose a medium for a reason.
“Text is the lowest bandwidth format. Interaction teaches ten times more.”
No evidence supplied establishes that universal order or multiplier. Mayer and Moreno describe benefits from relevant words and pictures together. Hegarty, Kriz, and Cate found no animation advantage over static diagrams in their mechanical-system experiments. Design and task matter.
Mayer & Moreno, 2002 (PDF) ↗Hegarty, Kriz & Cate, 2003 (PDF) ↗“The marginal cost of generated code is essentially zero.”
Code generation uses computation and may incur usage charges. Verification, debugging, integration, and maintenance still take work. Small tools can be economical without being free. Custom scripts and one-off software also existed before generative AI.
Example primary source: model usage pricing ↗The Karpathy quote about discardable artifacts.
Secondary reproductions attribute the idea to Andrej Karpathy. The original post could not be retrieved in this review. We therefore do not present its wording as a verified quotation, or its format preference as experimental evidence.
Reported original post; content not independently verified ↗“A learning rate of 1.5 causes divergence.”
That depends on the objective. For f(x) = x², xₙ₊₁ = (1 − 2η)xₙ. At η = 1.5, each step multiplies x by −2. The magnitude grows. For f(x) = ½x², that same learning rate would converge.
“Thirty seconds to generate. Ten seconds to understand.”
Those are personal timings, not reproducible promises. They do not establish a general learning speed or a reliable time to generate an explanation.
03HUMANIZER SOURCE REVIEW · CHECKED 20 OCT 2026
Documented behavior, not a contest.
Features are based on project documentation and the supplied listings. No comparative writing test was run.
| Listing | What was checked | Limits |
|---|---|---|
| Humanizer · blader ↗ | The README documents sample-guided voice matching and preservation of source facts. | The supplied Oct 1 directory date was not independently confirmed. |
| Matematic Humanizer ↗ | The project documents English, Polish, and Portuguese editions, plus checks after editing. | The supplied v1.2.0 and Oct 18 directory metadata were not independently confirmed. |
| Humanize Kit ↗ | The README documents voice-profiler, humanize, humanize-ig, sample-based profile creation, and MIT licensing. | The supplied v2.2.0 and Sep 29 directory metadata were not independently confirmed. Detector-related claims are not used as evidence of writing quality. |
| humanizer · adelaidasofia ↗ | The screenshot identifies the repository. Its SKILL.md confirms v3.0.0, a local voice indexer, and English/Spanish rules. | Fingerprint setup was reviewed in source, not executed. The supplied Sep 29 directory date remains unconfirmed. |
See GitHub counts for all reviewed projects ↓. Some official STE pages blocked direct retrieval; relevant passages were available through indexed extracts. This site is not affiliated with ASD, Anthropic, the plugin authors, or Andrej Karpathy.
Choose an editing approach ↗04CLAUDE SKILLS
Reusable instructions, with a defined scope.
Skills can replace repeated setup.
Anthropic documents automatic selection of relevant skills and explicit invocation. A skill supplies instructions and resources; the user still needs to provide a task and its inputs. Automatic invocation can be restricted in Claude Code.
Claude Code skill documentation ↗Claude app setup ↗Two Stayux listings; one current project.
The old STAYUX-STE-LINTER README redirects to Stayux-Clear-Answer. Writing Skills includes an ASD-STE100 skill. STE Explain supplies two Markdown instruction files. HasData handles web data. All four repositories were retrieved for the source review on 21 October 2026.
Source review and repository links ↗All reviewed skills and plugins.
| Skill or plugin | What it does | GitHub stars |
|---|---|---|
| Humanizer · blader ↗ | Natural rewriting and voice matching | ★ 59,735 |
| Writing Skills ↗ | Writing frameworks, including ASD-STE100 | ★ 49 |
| Stayux Clear Answer ↗Formerly STAYUX-STE-LINTER | Plain English and a technical-writing linter | ★ 3 |
| STE Explain ↗ | Simplified Technical English explanations | ★ 3 |
| humanizer · adelaidasofia ↗ | Voice profiling from your writing | ★ 2 |
| Humanize Kit ↗ | Voice profiling, rewrites, and Instagram copy | ★ 2 |
| HasData ↗ | Web data collection | ★ 1 |
| Matematic Humanizer ↗ | Editing in English, Polish, and Portuguese | ★ 0 |
Use CLAUDE.md for recurring rules.
A concise CLAUDE.md can cover everyday writing guidance without an extra skill. Add a skill for task-specific workflows, scripts, or reference material. This is our recommendation, based on Claude Code’s documented distinction between standing context and on-demand skills. Instructions still need review; they are not enforced security controls.
Claude Code instruction guidance ↗Review the behavior, not the star count.
Static review on 21 October 2026 covered manifests, relevant skill instructions, local script calls, hooks, and MCP configuration. No obvious malicious behavior was found in the reviewed writing paths. We did not install the plugins, run their code, audit remote services, or test release binaries. This is not a security certification.
Findings and exact source versions
Stayux Clear Answer: The Python linter uses the standard library and reads stdin plus optional local dictionaries. No network or subprocess calls were found in that linter. The duplicate checker reads local skill headers; /init edits CLAUDE.md. Pass untrusted text through a file or stdin without interpolating it into shell commands. Reviewed source: bebc8ff ↗
Writing Skills: The writing skills are Markdown instructions. No runtime executable scripts, hooks, MCP configuration, or credential requests were found in the reviewed skill package. Repository validation and publishing tools are separate. Reviewed source: 2e692b8 ↗
STE Explain: Its skill folder contains two Markdown files with no runtime code or dependencies. The repository also has development scripts and a local validation hook; the skill folder does not include them. Reviewed source: f2dd1d8 ↗
HasData: Configures a remote MCP connection to mcp.hasdata.com. Queries, URLs, and any supplied request data leave Claude for that service. Its session hook prints routing instructions. It includes prompt-injection and credential-handling guidance, but these are instructions, not enforced protection. Optional CLI installation introduces a separate binary that we did not audit. CLI examples place keys in command arguments or environment variables; prefer interactive configuration and avoid secrets in shell history. Reviewed source: 0423a81 ↗
These findings apply to the linked source snapshots. Recheck changes before updating. We use no minimum-star safety threshold.
CLAUDE.md provides standing context.
Our concise starter is written for coding in Claude Code: scoped changes, tests, and checks it must actually run. From the reviewed Clear Answer, STE Explain, and Writing Skills sources, its writing section keeps consistent terms, direct steps, conditions before actions, and preserved qualifications. It omits their prose-linting loops and strict vocabulary rules. This is editorial guidance, not tested model performance or STE compliance.
Official file and loading guidance ↗Preview and download CLAUDE.md ↗