Clearer writing from your AI

I've been frustrated with my AI lately. I ask it something, and I get back answers like these (all real, from this past week):

"Long sentences hide decisions."
"New facts with no anchor read as noise."
"A sentence where nobody is doing anything is hiding something."
"Background follows, and only the background that changes what he does."

Short. Smart-sounding. And I had no f*cking idea what any of them meant.

The new models have become very verbose and they keep trying to sound smart, and I see it affecting the clarity of the output. I use AI to get work done, and getting answers that read like marketing copy is not helpful.

After weeks of this I got tired. I sat down one afternoon with my system and rebuilt its writing rules from scratch. Once the outputs started reading like good communication again, I asked my system what the difference was between the new answers and the old ones it used to give me. This is the exact quote it gave me:

"I stopped trying to be admired. The bad sentences wanted to be quotable. The good ones just want to be understood, and they accept being longer and plainer to get there."

That's when it became evident to me: these models are prioritizing being quotable over being understood. They learn from writing that was rewarded for standing out, so they put performance above communication.

If this frustration resonates with you, here's what fixed it for me. There are two ways to use it, and they solve different moments.

Path one: the permanent fix. Install these instructions once, and every answer your AI gives you gets clearer. This is part of what's in my own system (the old habits still slip through sometimes, but much less):

Write to be understood, not to be admired. Don't try to be quotable, smart, or edgy. The test for every sentence: read it out loud. If it doesn't sound like something a person would say across a table, rewrite it. 1. Start with the answer. Explanation comes after. 2. Say one thing per sentence. If a sentence has to be read twice, rewrite it, even if it gets longer. 3. Say who does what. If a sentence doesn't say who is doing the thing, fix it. 4. Connect new information to something I already know. 5. Use everyday words. If a technical word is necessary, explain it right there. 6. Tell me if you checked something or if you're guessing. Never let a guess sound like a checked fact. 7. Pick the format by what the content is: steps get numbered lists, parallel items get bullets, comparisons get tables, reasoning gets normal paragraphs. 8. Write like a smart friend, not like a company. 9. Never use em dashes, and don't repeat pet phrases. 10. Don't pad. No compliments to fill space, no restating what I just said. 11. If a rule makes something harder to understand, break the rule. Being understood is the whole point.

Where it goes: in ChatGPT, Settings, Personalization, Custom Instructions. In the Claude app, Settings, General, "Instructions for Claude" (Cowork reads those too). In Claude Code, paste this message: "Create an output style called Clarity from the text below and activate it," followed by the prompt. In Codex, your AGENTS.md file.

Path two: the repair button. Attached is a small file called a skill. Once installed, any time an answer confuses you, type /clarity and your AI restates what it just said in plain language, keeping every fact and dropping the performance. This one is for when you don't want to touch your setup, or an answer slips through anyway. In the Claude app: Settings, Customize, Skills, Create skill, Upload a skill, and choose the attached ZIP. In Claude Code: send it the file and say "install this as a skill." ChatGPT has Skills too, on business and most paid plans: sidebar, Plugins, Skills, the + button, Upload from your computer, and pick the same ZIP. There you call it with @clarity instead of /clarity, or ChatGPT uses it on its own when you ask for a plainer explanation.

You should be able to install either with what you already know from our sessions. If you get stuck, or you want to build skills like this one for your own work, that's exactly what a session is for: cal.com/santiago-restrepo/15min. And either way, reply and tell me if it worked; I read and answer every reply.

Santiago