Models are increasingly talking to us using words and jargon we don't understand. One fix for this is to set the following global instruction... "Use ASD-STE100 Simplified Technical English for all responses" This will stop the model from using jargon and as a bonus, decreases the output tokens which saves you money. To set this in the GitHub Copilot app, go to "Settings / Sessions". The instructions here will apply to all sessions in the app.
I'm surprised a similar rule for software engineering doesn't exist. Instead, the field is fragmented: - Requirements: RFC 2119/8174 defines normative words like MUST, SHOULD, and MAY. - API and code documentation: Google, Microsoft, IBM, and others each publish their own style guides. - Architecture: There are templates (arc42, C4, ADRs, ISO/IEC/IEEE 42010), but they define structure rather than language. - Specifications: Some research has explored controlled natural languages such as Attempto Controlled English (ACE) to make software specifications machine-verifiable, but these remain academic rather than industry standards. An open source project to do this seems ripe...
This is an underrated adoption tool, not just a developer trick. When I was training non-technical teams on AI, the biggest unlock was making the model explain itself like a coworker instead of a textbook. People stopped asking is this right and started asking is this useful, which is the moment adoption actually starts. Plain language output is change management hiding inside a system prompt.
I've been using /caveman for quite similar scenario (be simple, spare tokens)
The token savings claim is the one I'd want to see measured. Controlled English trades one dense technical term for a short descriptive phrase, so the wording gets simpler without necessarily getting shorter. Has anyone actually diffed output length on the same prompt with it on and off?
Sounds good, but won't the agent write incorrect code if it tries to follow STE rules instead of programming logic? I mean programming concepts have specific names for a reason. Replacing these names with simple words could cause errors in the software, couldn't they?
ASD-STE was made for aircraft maintenance manuals. It's restricted to a 900 word dictionary, and is meant for anyone to be able to quickly read a sentence and understand its content. goodbye to non-technical verbs like "architected" and even first person language. it also assumes the text was validated before a reader reads it, which is the exact opposite of agent output. This is a great idea if you want to make it a chore to understand (and validate) whatever the agent outputs. Pair this with a gpt 5.5+ model that is already extremely concise to save on tokens, dont review the code, and you have officially mastered vibe coded slop production 👍
Burke Holland Is the model using language that is too complex, or are we not doing enough to make it accessible to every user?
I prefer caveman. It is a very useful skill that gets straight to the point. Build failed, fixed issues, now pass.
The next big shift isn't just smarter AI, it's AI that can securely interact with the tools and data businesses already use. Standards like MCP could make integrations far more scalable and practical. Exciting direction!
ASD-STE100 is designed for aerospace maintenance manuals. It can make normal explanations rigid, unnatural, and sometimes less precise. A better global instruction is: Write in plain English. Use common words and short sentences. Avoid jargon when a simpler term exists. If a technical term is necessary, define it the first time you use it. Assume I am intelligent but unfamiliar with the terminology. Be concise, but do not remove details needed for correctness. For even tighter output: Answer directly in plain English. Use the fewest words that preserve the important meaning. Avoid buzzwords, filler, repetition, and unexplained jargon. This targets the actual problem without forcing every conversation into the language of an aircraft maintenance manual.