Your Old AI Prompts Are Costing You Money (9 Fixes)
📓 The prompt pack — all nine as copy-and-paste blocks, free, no email. · 🧭 Deciding whether to pay for 5.1 at all? That's the other half: Claude Fable 5.1 for Creators — Care or Skip? · 📄 The source, so you can check me on any of it.
Same prompt. The good one — the one you've pasted into every project for a year, the one you'd be genuinely annoyed to lose.
You run it this morning and what comes back is... fine. Flatter. Shorter. It stopped pushing back on you. It stopped asking the one follow-up question that made it worth using.
So you run it again. Same flat thing. And you sit there for a second wondering what you did wrong.
You didn't do anything wrong. Your prompt didn't break. It expired.
What actually changed
Anthropic published a page on what changed in Claude Fable 5.1. It's forty screens of engineering notes, so almost nobody outside a dev team has read it.
Their own first line is that your old prompts should still work fine. And they will. Fine is the problem. Fine is not what you were getting a year ago, and there are nine specific reasons why.
Seven of these are copy-and-paste. The other two are a setting and a habit. None of it is code, and the whole thing takes about ten minutes.
Here's why this hits harder when you're solo. You don't have a prompt engineer. Your instructions file is the whole company handbook — written by you, in a hurry, on advice that was true at the time, and nobody's maintaining it. So it rots. And you feel that as "the AI got worse."
If you only use this a few times a week, fixes 2 and 4 are the ones you'll feel, plus fix 5 if you ever paste an article in. If it's running your admin, all nine are yours.
Part one: what to change in your settings
1. Effort is your bill, not a quality slider
Fable 5.1 has five effort levels: low, medium, high, extra high, and max. High is the default and it's the right place to start.
One caveat before you go looking: you only get this dial if your setup exposes it — Claude Code, the developer API, most builder tools. In a plain chat window it isn't yours yet.
Here's what matters if you're paying your own bill. At medium, Anthropic says results "roughly match Claude Fable 5 at lower cost." At low, it's "often competitive with Claude Opus and Claude Sonnet models on cost per task while scoring higher."
Read that again if you've been defaulting everything to the top setting. Cheaper and better, for the right kind of work.
So batch-writing captions, cleaning up transcripts, renaming files, sorting an inbox — that's low-effort work. Literally. Run it there. Save the expensive settings for the thing that actually needs thinking: your strategy, your positioning, your pricing.
The move is not "always max." The move is the cheapest setting where your work still holds up. Always max is a bill you're writing yourself.
2. Delete your expired rules
This is the one I'd do first if you only do one.
Open your instructions right now. Custom instructions, project files, that block of text you've been dragging from tool to tool since 2024.
Every line in there that says "stop using bullet points," or "don't over-format," or "hold your findings until the end" — those were written for an older model that overdid it.
This one already underdoes it. Anthropic's own wording: 5.1 "uses bold less and is less likely to reach for headers, lists, or quotation marks."
You are pushing on a door that's already closed. And you're paying tokens to push.
I did this on my own file before I made the video. Four dead rules. One of them I'd been dragging around since GPT-4 and I could not tell you why. Took me nine minutes.
Deleting is only half the documented fix, though. If you want something in that slot, don't write "less formatting" — write when formatting is right. That exact block is in the pack.
3. Ask it for progress updates
If you ever hand it something big and walk away — a whole content calendar, a full site audit, a month of bookkeeping — this model says a lot less while it's working than the last one did. You hand it over and you watch nothing happen for minutes.
That's not it stalling. That's it working.
But if you're handing it a twenty-minute job, you want to see it thinking, so you can catch it going the wrong direction at minute two instead of minute twenty.
You just have to ask: say what you're about to do, give me short updates while you work, and end with a recap that stands on its own.
That last part is the bit you'll actually use. A standalone recap is what you forward to a client, or paste into your notes, without rereading the whole thread.
Part two: what it's doing to your content
4. Kill the mannered prose
The writing is better than it used to be. Genuinely. But it can get dense — long sentences, big walls, a little pleased with itself.
You know the sound. Nothing is ever "a setting worth adjusting," it's "a dial worth turning." Nothing "matters," it "earns its keep."
There's a phrase for this straight out of the docs, and most of the time it's all it takes:
Please remove all mannered prose.
Mannered means writing that performs instead of saying the thing. Yes — including the door line I used two fixes ago. That one's paying for itself. Most of them don't.
Put that line in your instructions once and your captions, your emails and your proposals all get shorter and more direct. Permanently.
5. Check the quoting before you publish ⚠️
This is the one that matters most and gets talked about least.
Anthropic documents it plainly: 5.1 "is more likely than Claude Fable 5 to reproduce passages of the source text without marking them as quotations."
Because that's most of us. If you're feeding it articles and getting back scripts, newsletters, carousels, posts, client reports — you might be publishing somebody else's sentences under your name and not know it.
Anthropic documents the behaviour. What it costs you is my read: if a publisher recognises their own sentences in your video, that's a takedown request. With a client, it's a much shorter conversation than that.
Either way it is not the model's problem. It's yours. Your name is on the upload.
Two things. One: ask for it in its own words, explicitly, every time. Two: anything that comes back sounding a little too polished, a little too finished — paste one sentence of it into a search engine before it goes out.
Ten seconds. Every time.
6. Let it zoom in
It reads images better now. But on the dense stuff — your analytics screenshots, a spreadsheet, a contract, a dashboard with forty numbers on it — it does its best work when it can crop into one region and enlarge it, rather than reading the whole frame at once.
So crop to the part you actually care about before you paste it. And if your setup can run code on the image, let it crop and zoom itself — that's where the docs say the real gain is.
It's the difference between reading one number and squinting at forty.
Part three: making it actually do the work
7. Make it look things up
At low effort, it's more likely to answer from memory than to actually look things up.
Memory is a real problem in this space, because everything moves monthly. Prices, model names, what's in which plan, which tool just got acquired.
So anything about current pricing, a competitor, a tool that changed last month, a platform's rules — turn the effort up for that question, or tell it flat out: search the name exactly as I wrote it before you answer.
Half-remembering something is exactly what makes a stale answer sound confident. And a wrong number in a client proposal costs you a lot more than the ten seconds it takes to check it.
8. Tell it to stay in its lane
It will do more than you asked. Anthropic documents this on code, but you'll recognise it anywhere. You ask for one email and get a whole sequence. You ask it to fix a headline and it rewrites your landing page.
That sounds generous. It isn't, when you're the one who has to read all of it and decide what to keep.
So tell it: if you notice something else worth doing, put it in the summary as a suggestion. Don't build it.
You get your one thing, done properly, plus a list of what it spotted. That list is genuinely useful. It's just not the deliverable.
9. Make it actually finish
And this is the exact opposite problem.
On long jobs it'll sometimes tell you what it's about to do instead of doing it. Or it stops to ask permission for something you already asked for.
Fine if you're sitting right there. Not fine if you kicked off a job, went to go shoot or into a client call, and came back an hour later to a message asking whether you'd like it to proceed.
Two instructions fix this. First: you're operating autonomously, I'm not watching, so don't ask — if it's reversible and I already asked for it, just do it.
Then: before you end your turn, look at your last paragraph. If it's a plan, or a question, or a promise, go do that work right now.
That second one is the difference between coming back to a finished draft, and coming back to a really well-written description of a draft.
One trade-off worth knowing: tell it not to ask, and it also stops checking in on the genuinely ambiguous stuff. That's a fair trade for me. Might not be for you.
The catch
None of this makes the output good.
It makes it accurate, and it makes it cheaper, and it makes it finish. That's it. The taste is still yours. Every one of these just gets the machine out of your way faster.
That's the actual deal. And anybody selling you something else is selling you something.
The nine, in one line
Cheapest effort that still holds · delete your expired rules · ask for progress updates · kill the mannered prose · check the quoting before you publish · let it zoom in · make it search · tell it to stay in scope · tell it to finish.
None of that is developer work. That's just knowing how to talk to the thing you're already paying for.
If you only do one of them, do number two. Open your instructions tonight and read them like somebody else wrote them. I'd bet money there's a line in there costing you output quality and tokens right now.
📓 Get the prompt pack — all nine as copy-and-paste blocks plus the ten-minute setup. Free, no email.
🧭 Still deciding whether 5.1 is worth paying for? This piece assumes you're already using it. The other half of the question — what actually reaches you, what doesn't, and a one-week test — is in Claude Fable 5.1 for Creators: Care or Skip?
The behaviours above came from Anthropic's own documentation. What they mean for your business is my read, not theirs.
AI handles the busywork. You keep the taste.