Teach your AI that sometimes less is more.

Claude Opus 5 and Sonnet 5 are the most capable models we have used. They are also, out of the box, exhausting to read. Long, hedged, self-analytical, and fond of a metaphor nobody asked for.

That is a usability problem before it is anything else. If you have to read four paragraphs to find the one line you needed, the tool is costing you time rather than saving it.

We have written about two fixes for this recently. Both are copy and paste. Here they are in one place, with the exact text.

Fix one: one sentence, globally applied

This is the simplest approach we have seen. Add a global agent rule:

ASD-STE100 is a controlled English standard written for aerospace maintenance manuals. Restricted vocabulary, short sentences, one meaning per word. It exists so an engineer at 3am cannot misread an instruction. Point a chat model at it and you get the answer without the performance.

The reason it works is worth stating plainly, and here is how we put it on LinkedIn:

There is a side benefit. Fewer tokens out means less compute, which means a lower energy cost per answer. That is not the main reason to do it, but it is a real one.

Where the rule goes

• If you use Claude Code or Cowork, put it in your global CLAUDE.md.
• If you use Codex, Cursor or anything else that reads AGENTS.md, put it there.
• If you only use the chat apps, open settings in Claude or ChatGPT and add it to your global or custom instructions.

Same sentence in all three places. It works for anything AI-related.

Fix two: the Anti-Slop Directives

The ASD-STE100 rule handles length. It does not always handle tone. For that we use a longer block, and the reason for it is simple:

This goes in the same three places as the rule above. Copy, paste, done.

1. Ban the Victorian Academic Tone & Pseudo-Profundity

• Never describe mundane logic, variables, or architecture as load-bearing, foundational tapestry, epistemic tension, delve, or intricate dance.
• State the technical reality directly. Do not romanticize simple engineering concepts with architectural metaphors.

2. Strip Paraprosdokians and Melodramatic Antitheses

• Strictly forbid tragic setup/reversal copy constructions (e.g., "Ten features. Zero headaches." or "Everything about billing changed. Your invoice didn't.").
• Write straightforward technical and marketing copy. State what the tool does, how to use it, and what it integrates with without dramatic flair.

3. Prohibit Unsolicited Refactoring ("I've Taken the Liberty")

• Execute strictly what is requested. Never refactor surrounding architecture, create modular layers, introduce new dependencies, or rewrite files unprompted.
• If a 1-line fix is requested, return only the 1-line fix.

4. Eliminate Sycophantic Apologies and Epistemic Hedging

• When corrected, acknowledge the fix in one brief clause (e.g., "Corrected.") and provide the accurate output.
• Ban self-flagellating over-apologies (e.g., "You're entirely right, thank you for your sharp eye, let me recalibrate").
• Do not preface standard disagreements or technical tradeoffs with "I want to push back slightly" or "To be candid rather than merely encouraging."

5. Enforce Direct Structural Openings (Zero Fluff)

• Do not generate introductory meta-announcements (e.g., "Here is the solution...", "Let's examine the fascinating tension between...").
• Lead immediately with the code, command, or data.

Directive 3 is the one that surprises people. Unsolicited refactoring is not a tone problem, it is a scope problem, and banning it in the same block turns out to be the most useful line of the five.

And come on, what a word 'Paraprosdokians' is.

What it actually changes

Outputs get shorter, so they are quicker to read and cheaper to run. Corrections get acknowledged in a clause instead of a paragraph of apology. Answers start with the answer.

Two things to watch for. The model gets blunter, and some people find that cold. And if you are using AI for creative writing rather than work, these rules will flatten it, so keep a separate profile or turn them off.

Neither of those has been enough to make us take the rules out. We have had them on for months.

Try it for a week

Add the one-line rule first. Give it a few days. If the tone still grates, add the directives underneath it.

If you cannot find where your global instructions live, or you want help setting this up properly across a team, book a discovery call or drop us a message. It is a ten minute job and it makes the next thousand prompts better.