Anthropic Shares AI Prompts

Anthropic has published a series of prompting techniques designed to help users get better results from Claude while using fewer tokens and making more effective use of lower-effort settings.

Although the recommendations were developed around Claude Fable 5.1, several of the ideas could also be useful when working with other AI models.

Make AI Writing More Direct

One of Anthropic’s suggestions focuses on a common problem with AI-generated writing: unnecessarily elaborate language.

The company refers to this as “mannered prose”, where a simple point is replaced with metaphors, flourishes or phrases that sound more like an attempt to impress than communicate information.

Rather than simply telling an AI model to “write better”, Anthropic recommends clearly explaining the type of language to avoid and what should replace it.

For example, users can explain that when a straightforward phrase is available, the model should use it instead of introducing unnecessary metaphors or decorative language.

Anthropic also suggests that the instruction can be kept extremely simple:

“Please remove all mannered prose.”

The idea is that defining the unwanted writing style can help produce clearer and more natural responses without requiring a long list of individual rules.

Review Old Formatting Instructions

Anthropic has also highlighted changes in how newer Claude models handle formatting.

Fable 5.1 reportedly uses less bold text, fewer headings and lists, and makes less use of quotation marks than earlier versions.

This could create problems for users who have older prompts containing strict formatting instructions.

Anthropic recommends reviewing these instructions rather than assuming that a prompt written for an older model will continue to produce the same results.

A more flexible approach is to tell the model when formatting is useful. Lists and bullet points can be used when information is complex, while straightforward conversations can remain in normal prose.

Users who specifically request minimal formatting should have that preference take priority.

The broader lesson is that prompts may need to evolve alongside the AI models they are being used with.

Getting Better Results From Lower Effort

Claude Fable 5.1 includes different effort settings, allowing users to balance performance against resource use.

Anthropic recommends starting with the higher setting before testing whether lower levels can produce an acceptable result. According to the company, the low-effort setting can still perform strongly compared with some more expensive models.

There is a potential drawback, however. Lower-effort responses are less likely to make use of search and retrieval tools and may rely more heavily on the model’s existing knowledge.

That can become a problem when a question involves information that changes quickly.

Anthropic therefore suggests using a prompt that encourages the model to verify unfamiliar or rapidly changing information through search.

This is particularly useful for subjects such as AI models and developer tools, where information can become outdated within a matter of months.

The instruction effectively tells Claude that recognising a name does not necessarily mean it should trust its existing knowledge. If the subject is unfamiliar or fast-moving, the model should search first and include the exact name supplied by the user in its search.

This can help reduce the risk of an AI giving a confident answer based on outdated information.

Make Smaller Changes When Editing Code

Another recommendation is aimed at developers and anyone using AI coding assistants.

Rather than rewriting an entire file when only a small change is required, Anthropic recommends making the most targeted edit possible.

The reasoning is straightforward: unnecessary rewrites use more tokens and can introduce changes that were not part of the original request.

A simple instruction can encourage the model to edit only the relevant section:

“When it will not affect the end result, try to surgically edit a file rather than rewrite the entire thing.”

This approach could be particularly useful when working with large CSS, PHP or JavaScript files, where a complete rewrite creates more opportunities for unintended changes.

Tell AI to Finish the Task

Anthropic has also identified situations where AI agents stop before completing a task.

This can happen when an agent is working through a long or asynchronous process and pauses to ask whether the user wants it to continue.

For tasks where the user’s original request already provides enough authority to proceed, Anthropic recommends explicitly telling the model to work autonomously.

The instruction can establish that the AI should continue through reversible steps without repeatedly asking for confirmation. It should only stop when it reaches a destructive action or a genuine change in scope that requires a decision from the user.

There is an important exception. If someone is simply asking a question, describing a problem or thinking through an issue, the AI should provide its assessment rather than automatically making changes.

Anthropic also recommends adding a final check before the model ends its response.

If the last part of the response is simply a plan, proposed next step or promise to carry out additional work, the model should complete that work instead of stopping prematurely.

This can help prevent AI agents from responding with what they intend to do rather than actually doing it.

Check Before Making Changes

The same instruction also includes a safeguard for actions that alter system settings.

Before running commands that could restart systems, delete files or modify configurations, the AI should ensure that the available evidence genuinely supports the action.

This is particularly important for AI agents because a pattern that resembles a familiar technical problem does not necessarily mean the underlying cause is the same.

The model should therefore avoid making potentially disruptive changes based solely on an assumption.

Why These Prompting Techniques Matter

Anthropic’s recommendations highlight a broader point about working with modern AI systems.

Better results do not always require longer prompts. In some cases, a short instruction that clearly defines an unwanted behaviour can be more effective than a large collection of detailed rules.

The recommendations also show why old prompts can become less effective as models change.

Formatting preferences, tool use, coding behaviour and the way models handle long-running tasks can all evolve between versions. Instructions that worked well previously may therefore need to be reviewed and simplified.

For businesses and individuals using AI regularly, this makes prompt maintenance increasingly important.

Key Takeaways for AI Users

Anthropic’s latest guidance offers several practical lessons:

Be specific about poor writing habits. Tell the model which styles or phrases you want to avoid rather than simply asking for better writing.
Keep formatting instructions flexible. Explain when headings, lists or bold text are useful instead of imposing them on every response.
Verify rapidly changing information. Lower-effort AI settings may rely more heavily on existing model knowledge, so prompts can encourage searches when information may be outdated.
Use targeted code edits. Ask AI to modify only what is necessary rather than rewriting entire files.
Tell agents when to continue. For autonomous tasks, make clear that the model should complete the work rather than repeatedly ask for permission.
Keep safeguards in place. Potentially destructive system changes should only be made when the evidence supports them.

Anthropic’s prompting guidance is specifically intended to improve performance with its latest Claude model, but many of the principles are relevant to AI tools more generally.

As models become more capable, the focus is shifting away from simply writing longer prompts towards giving AI systems clearer instructions about when to search, what to change, when to continue and what behaviour to avoid.