How to Make ChatGPT and Claude Stop Burying the Answer

You asked for one thing. You got an essay. The answer is probably in there — third paragraph, maybe — but by the time you've scrolled past the throat-clearing and the "great question!", you've half forgotten what you asked. So you skim, skip back, skim again — and miss the part that mattered.

If your brain loses the thread easily — and plenty do, with or without an ADHD diagnosis — that wall of text isn't just annoying, it's really annoying. Start five chats at once and it has every chance of driving you up the wall. The good news: the AI isn't doing it on purpose, and you can switch most of it off in about two minutes.

The skill

Claude Code only

Cost

Free, MIT

Fix it yourself

2 min in settings

Works in

ChatGPT · Claude

Why your AI keeps burying the answer

A model doesn't know how much you want. Faced with that, it defaults to thorough, because thorough looks safe and complete looks helpful. So it restates your question, lists every caveat, walks you through its reasoning, and closes with an offer to keep going. Nobody asked for any of that. And, honestly, sometimes it sounds like it just enjoys the sound of its own voice.

For a reader who skims by default, "complete" is the opposite of helpful. The thing you needed is now competing with six things you didn't. You're not lazy for wanting the point first — you're outnumbered.

The AI isn't trying to bury you. It defaults to thorough because thorough is safe — so you have to tell it not to be.

The skill that promises to stop it

Someone built a fix and gave it the most on-the-nose name possible: i-have-adhd. It's a free, open-source skill for Claude Code — Anthropic's terminal tool for developers — written by an independent developer, not by Anthropic. It's ten rules that reshape how the AI answers: lead with the action, number multi-step work, no preamble, no "great question", no "hope this helps", matter-of-fact when something breaks, cap lists at five, and say where you are in a long task.

This is exactly why it spoke to me. Half of last week was me nudging Claude with the same questions — where are we again? what's next? just give me today's three that matter? — often enough that a skill built to answer them upfront felt aimed straight at me.

Read plainly, it does exactly what it claims and nothing sneaky. It's not a program that runs on your computer — it's a page of instructions your AI reads. In Claude Code it stays off until you type /i-have-adhd, so it can't quietly take over. As output-shaping goes, it's clean and well made.

Two catches, though. First, it lives in Claude Code — the developer tool. If you're in ChatGPT or the normal Claude chat, you can't install it at all. Second, a smaller thing worth knowing: the version for a different tool (Codex) can switch itself on even though the docs promise it stays off until you ask. Minor, but it's the kind of gap you only find by looking.

There's a quieter cost, too: if you only tell the AI to tighten up after it's dumped the wall of text, you've already paid — in tokens, in waiting — for output you never wanted.

The two-minute fix, in any chat

Here's the twist: you don't need the skill to get what it does. Both ChatGPT and Claude already carry a settings box that quietly rides along on every conversation — ChatGPT calls it Custom Instructions, Claude keeps it in your Profile, and inside a Project you can set it per workspace. Whatever you put there, the model reads before it answers you. Every chat. No repeating yourself.

Paste it in and you've told it, once, how to talk to you:

✦ Paste into your settings ChatGPT · Claude Copy

How to talk to me:

• Lead with the answer. First line = the thing I can act on (the step, the decision, the command), not a recap of my question.
• No preamble, no closer. Skip "Great question", "Sure!", "Hope this helps", "Let me know if...". Start at the answer, stop when it's done.
• Don't flatter me. Drop "great idea", "you're right", "excellent point". If I'm wrong or there's a simpler way, say that first.
• Number any multi-step answer. One action per step.
• Keep it short. One line if one line does it. If a list runs past five items, split it into "now" and "later".
• On a longer task, tell me where we are: "Step 3 of 5 done. Next: X."
• Real estimates, not vibes: "about 10 minutes", not "a bit of work".

Change the lines you don't like — that's the point of it being yours. Want more depth on hard questions? Add "when I say explain, go long." Prefer bullet points to numbers? Say so. The block is a starting shape, not scripture.

The part no skill does

Look at line three: don't flatter me. That's the one that does the heavy lifting, and it's the one the skill leaves out. Making an AI shorter is only half the job. If half of what's left is "great idea" and "you're absolutely right", you've traded a long answer for a padded one — the fluff just moved to the front.

Flattery is the sneakiest kind of filler, because it reads as friendly. But you didn't come for a cheerleader. You came for the answer, and ideally for the part where it tells you your plan has a hole in it before you build on it. Telling the model to push back instead of agreeing is worth more than any "be concise" you'll ever write.

Brevity isn't enough if half of what's left is "great idea". Flattery is just more text between you and the answer.

One honest caveat: instructions like these leak. Over a long chat the "great question" creeps back in, and now and then you'll have to nudge it — "shorter", "drop the preamble" — to bring it back into line. That's normal. Tuning how your AI talks to you is a habit, not a switch you flip once. It still beats reading five paragraphs a hundred times a week.

So do you need the skill?

If you live in Claude Code, drowning in verbose output, and you haven't set any of this up — sure, /i-have-adhd is a fine shortcut. It's free, it's clean, and it saves you writing your own rules. Install it and move on.

For everyone else — in ChatGPT or Claude's normal chat — skip it and paste the block. And notice what just happened: the "right" move was to not install the shiny thing. The reflex in the AI world is "looks good, lots of stars, download it." A calmer habit is to ask what problem you actually have, and whether a settings change solves it before a plugin does. Deciding against a tool isn't missing out. It's the plan part of chasing shiny objects with a plan.

Questions people actually ask

How do I make ChatGPT give shorter answers?

Custom instructions make it permanent: tell the model to lead with the answer, skip preamble and closers, and keep replies short unless you ask for depth. One setting, every future chat — no need to repeat it each time. For a single reply, just add "one line only" to your message.

Can I stop ChatGPT or Claude from being so complimentary?

Yes — add one line to your custom instructions: don't flatter me, drop "great idea" and "you're right", and if I'm wrong, say so first. The praise fades right away. It creeps back over very long chats, so an occasional nudge keeps it honest.

Are AI skills and plugins safe to install?

Sometimes — but "popular" is not "safe". A skill is instructions and, often, code that runs with your AI's access, so read what it does before installing. Most are harmless; the point is to look, not assume. Here's how to check one first.

The best answer is the one you'll actually read. ✦

That block sits in my own settings and still slips now and then — tuning never fully ends, and you don't need more than that. If you'd rather see what it looks like to take a skill apart piece by piece, the fable-method teardown does exactly that. And the CS Lab is where the tools worth keeping end up.