Anthropic's Power Prompts Unlock Better AI Results With Fewer Tokens

Anthropic has released a set of targeted prompting techniques for its Claude Fable 5.1 model that help users save costs, improve writing quality, and extract stronger performance from lower effort settings.

The guidance, published by Anthropic and reported by Search Engine Journal contributor Roger Montti on September 15, 2026, offers practical prompt upgrades that the company says may also benefit users of competing AI models — not just Claude. For teams still building familiarity with AI tools, understanding what artificial intelligence is and how it works provides essential context for getting the most from these techniques.

---

Cleaner Writing and Smarter Formatting

One of the most broadly applicable techniques targets what Anthropic calls "mannered prose" — the flowery, metaphor-heavy writing style that has become a recognisable AI signature. The company offers two versions of a prompt to address this problem.

The long version instructs the model directly: "Mannered prose substitutes metaphor and flourish for direct statement… The fix is to say what you mean. When a literal phrase is available, use it." For users who prefer brevity, the short version is simply: "Please remove all mannered prose."

This distinction matters because AI-generated content has increasingly drawn criticism for sounding performative rather than informative — a quality that erodes reader trust and undermines content credibility. The ability to produce clean, direct prose is now a measurable differentiator between AI outputs that convert and those that disengage.

Formatting Defaults Have Shifted

On formatting, Anthropic notes that Fable 5.1 uses fewer structural elements such as bold text, headers, and bullet points compared to earlier models. Users who previously built anti-formatting rules into their prompts may find those rules now conflict with the model's updated defaults.

Anthropic recommends either removing old formatting restrictions entirely or replacing them with context-aware rules. Its suggested prompt reads: "Use lists and bullet points when asked to, or when the content is multifaceted enough that they help with clarity. In conversational, personal, or emotional exchanges, keep to plain prose."

Periodically reviewing your formatting instructions is no longer optional — model defaults evolve, and prompts written against older behaviour can actively degrade output quality.

---

Getting More From Low-Effort Settings

Claude Fable 5.1 operates across multiple effort levels. Anthropic recommends starting at the default high setting and testing downward. The company notes that the "low" effort level is comparable in cost to Claude Opus and Claude Sonnet models while outperforming both on benchmarks — a meaningful cost advantage for teams running AI at scale.

However, a meaningful trade-off exists. At lower effort settings, Fable 5.1 relies more heavily on its parametric memory — what the model already knows — rather than actively searching for current information. In fast-moving fields like AI development, this can produce answers that sound authoritative but are quietly out of date.

The Search Nudge Prompt

Anthropic's nudge prompt addresses this directly: "When a query centres on a name you do not confidently recognise, or recognise from a fast-moving area like AI models and developer tools where the landscape shifts within months… search before answering… partial background is exactly what makes an out-of-date answer sound authoritative, so familiarity is not a reason to skip the search."

A useful parallel here: this mirrors the journalist's golden rule — when in doubt, verify before publishing. Familiarity with a subject is not a substitute for accuracy, and AI tools operating at reduced effort levels need an explicit instruction to prioritise currency over confidence.

For businesses weighing the business benefits of adopting artificial intelligence tools, cost-efficient performance at lower effort tiers represents a significant operational advantage — provided accuracy safeguards like this search nudge are built into standard prompts.

---

Surgical Edits and Autonomous Task Completion

Precision Editing for Developers

Two additional prompts target specific workflow pain points for technical users and power users working on longer projects.

For developers — including those working in WordPress, CSS, or JavaScript environments — Anthropic offers a targeted editing prompt designed to stop AI coding tools from rewriting entire files when only a small change is needed. The prompt reads: "When it will not affect the end result, try to surgically edit a file rather than rewrite the entire thing."

This approach conserves tokens and reduces the risk of unintended changes to functioning code — a concern that scales significantly in production environments where a misplaced rewrite can introduce regressions across interconnected components.

Autonomous Task Completion Without Unnecessary Interruptions

The final and most detailed prompt addresses a behaviour where Fable 5.1 unnecessarily pauses during long autonomous tasks to ask for permission before proceeding. Anthropic's solution instructs the model to act on reversible steps without interrupting the user and to stop only when facing "destructive actions or genuine scope changes."

The prompt also includes a self-check mechanism: before ending its turn, the model is told to review its last paragraph and complete any promised work rather than leaving it as an unfulfilled plan. This is a practical quality control layer built directly into the instruction set — one that meaningfully reduces the back-and-forth typically required to reach a complete output.

This level of autonomous operation reflects a broader industry shift toward AI agents capable of handling extended workflows with minimal human intervention — a development that carries both productivity benefits and the need for careful scope management. Organisations exploring how to operationalise these capabilities may find value in reviewing how small businesses are implementing artificial intelligence to streamline workflows without overextending resources.

---

What This Means for Everyday AI Users

Anthropic's prompting guide offers several immediately actionable takeaways.

Users frustrated by verbose or flowery AI writing can apply the mannered prose fix across most major models. Teams running AI tools at reduced cost settings can use the search nudge prompt to maintain accuracy without upgrading to a more expensive tier. Developers working with AI coding assistants can minimise costly and risky file rewrites by building the surgical editing instruction into their standard system prompts.

Treat Prompt Libraries as Living Documents

The broader lesson here is that AI model behaviour shifts with each new version. Prompts written six months ago may now produce suboptimal results simply because the underlying model has changed. Periodically auditing and refreshing prompt libraries is no longer optional — it is standard maintenance for anyone relying on AI tools in a professional setting.

The most effective AI users are not those with the most powerful models — they are those who understand how to instruct those models precisely, efficiently, and with an awareness of how defaults have changed. For a deeper perspective on how this technology continues to evolve, Anthropic's research publications offer primary-source insight into the decisions shaping model behaviour.