Claude Fable 5 Prompting Habits That Actually

Claude Fable 5 is real, it's out, and according to Anthropic, it's expensive — priced at $10 per million input tokens and $50 per million output tokens, which is double what Opus runs. That's the kind of number that turns "I'll just ask Claude" into a deliberate cost decision rather than a reflex.

Which is why the prompting conversation actually matters now. When a model is cheap enough that mistakes are noise, you can be sloppy. When it's priced like Fable 5, every wandering, underspecified prompt is a tiny invoice for your inattention.

Nate Herk, who runs the AI Automation channel and has clearly spent a lot of quality time with this model, put out a video distilling six prompting habits he pulled from his own experimentation, community feedback, and Anthropic's own documentation. It's a genuinely useful breakdown — and it surfaces something more interesting than a listicle: a coherent philosophy about what it actually means to communicate well with a highly capable model.

Give It the Why, Not Just the What

The first habit Herk emphasizes is context — and not just what you want done, but why you're doing it. His framing: "the context lets it connect your task to the right information instead of guessing what you meant."

The contrast he draws is clean. Don't say: "Write me an email to a client about the delay." Do say: here's the broader project, here's who the client is, here's what they need, now help me write that email.

This isn't a new idea — it's been good advice for every model going back years. But there's a reason it leads Anthropic's own prompting guidance for Fable 5, per Herk's reading of the docs: a more capable reasoning model doesn't just execute instructions better, it can actually use purpose as a constraint. Intent narrows the solution space. Without it, you're essentially asking the model to guess at your goals while optimizing for surface-level correctness.

What's interesting is the tension this creates with Herk's sixth habit: say less, not more. These aren't contradictory — give the why, don't pad the what — but it requires a kind of precision that most people aren't used to. Short but purposeful. Not terse, not verbose. It's harder than it sounds.

Negative Prompting: Tell It What Not to Do

Here's the intern analogy, which Herk leans on and I'll be honest, it works even though it's a little well-worn: "If you were explaining a task to an intern, you would tell them specific things to not do because they don't understand the process yet."

The example he gives is good enough to reproduce: instead of "take a look at this problem and handle it," try — "When I'm describing a problem or asking a question, the deliverable is your assessment. Report what you find and stop. Don't fix, send, edit, or delete anything until I say go."

That's a fundamentally different relationship with the model. You're not just describing a task; you're defining a boundary around what counts as completion. And Herk notes that negative prompting has gotten noticeably more effective as models have improved — a sign that reasoning capability and instruction-following are scaling together in ways that make constraint specification actually land.

The underlying mechanics: a language model is, at its core, predicting the most statistically likely next token. Left unconstrained, it will fill gaps with what usually comes next — which is often helpful, and sometimes deeply not what you wanted. Negative prompting is just being explicit about which defaults to override.

Don't Overplan — Let It Act

This one surprised me because it's a reversal of conventional wisdom. Herk says he's largely stopped using plan mode in Claude Code, a feature he used to recommend reflexively. His new approach: build your own logic for when the model should stop planning and start executing. The instruction he suggests — "when you have enough information to act, then act" — is almost embarrassingly simple, and probably effective for exactly that reason.

The context here is cost. On a pricier model, a long planning loop before any real work happens isn't just slow — it's burning tokens on deliberation that may not improve the output. Herk's suggestion to match effort level to task type is the same intuition: you don't need Fable 5 for everything. He puts it plainly — you probably only need to reach for Fable "5 to 15% of the time."

That's a striking claim and a useful one. The cost-performance tradeoffs in Anthropic's model lineup are real, and the temptation to default to the most capable model for everything is exactly the behavior that makes AI budgets spiral. Herk notes that per Anthropic's own comparison data, Fable 5 at lower effort settings can be competitive with Opus 4.8 at its ceiling — and cheaper. That's a meaningful signal about how to think about the stack, not just which model to use.

Make It Prove It

This is probably the most universally applicable habit Herk covers, and the one I'd argue matters most for anyone deploying AI in anything with real stakes.

The problem: models will tell you they're done when they're not. Or they're done, but they haven't verified anything. Herk's prescription is to bake verification into the instruction itself: "Before you tell me something is done, point to the result that proves it. Only report work you can show evidence for. If something isn't verified, say so plainly instead of guessing."

This is good regardless of which model you're using. But it's especially worth hardcoding into agent workflows, system prompts, and any CLAUDE.md files you're maintaining — because the instinct to treat a confident-sounding output as a finished output is a bias we bring to this, not something the model inflicts on us.

The Safety Routing Thing Is Kind of Wild

Okay, so this is where Herk surfaces something that I find genuinely strange — strange enough that it deserves careful handling rather than presentation as settled fact.

His claim: on Fable 5, keeping a standing instruction like "explain your reasoning" in your system prompt can trigger a safety check that silently routes your session to a less capable model without notifying you through the interface (though API users apparently do see an indicator). His read is that Fable 5 has guardrails built in that flag certain request patterns — particularly anything that looks like an attempt to surface the model's private reasoning process — as potentially suspicious, and routes those sessions accordingly.

I want to be clear: this is Herk's interpretation, not independently verified behavior. He's working from his own testing and his reading of Anthropic's internal logic. But even as a hypothesis, it's worth sitting with. The idea that asking a model to be more transparent about its reasoning could be interpreted as a red flag — and result in a less capable model answering your question, in silence — is a fascinating design tension. Transparency-seeking as a possible jailbreak vector. You can see the logic from a safety engineering standpoint, even as it produces a user experience that is... let's say, counterintuitive.

If Herk's read is right, this is the kind of architectural quirk that could produce real confusion in production environments where teams have standardized system prompts that include reasoning instructions. Flag it, test for it, don't assume it's documented behavior.

What All of This Actually Means

Here's my read, which I'll own rather than attribute: these six habits aren't really about Fable 5 specifically. They're about a shift in what the right mental model for interacting with capable AI looks like.

The old approach — dump context, be verbose, ask it to show its work, plan before acting — made sense when models needed hand-holding. The emerging approach is closer to working with a sharp colleague who gets annoyed when you over-explain and whose judgment you've learned to trust enough that you can give a direction and a boundary and get out of the way.

Herk's framing is that Fable 5 "follows short, clear direction better than older models do because it's just better at reasoning." That's the crux of it. Better reasoning means the model can infer more from less — which means the quality of your prompts becomes about precision of intent rather than volume of instruction. The advisor-model dynamics Anthropic has been building toward support exactly this: smarter models doing more with cleaner inputs.

The real question is whether users can make that shift — from treating AI as a system that needs micromanagement to treating it as something worth communicating clearly with. Fable 5's pricing is, in a weird way, a forcing function. When the meter is running at this rate, sloppy prompting has a cost you can see on an invoice.

That tends to focus the mind.

Yuki Okonkwo is Buzzrag's AI & Machine Learning Correspondent.