Stai usando l'effort di Claude nel modo sbagliato (ecco come sistemarlo)

Choosing Claude Code’s Effort Level

The video explains how to choose Claude Code’s effort level—low, medium, high, extra high, or max—instead of defaulting to the highest setting. Effort affects how deeply Claude investigates a task, how many checks it performs, and how many decisions it makes before asking for guidance.

The model determines Claude’s underlying capability. A stronger model at lower effort may outperform a less capable model given more time.

When to Use Each Effort Level

• Low: Use for quick drafts and tasks where you want to guide Claude step by step. Examples include brainstorming a video outline, shortening an email, finding a newsletter subject line, writing a short summary, or classifying support requests using clear rules. Low effort can also work for small revisions to material developed in the same chat. The video notes that, with some models, changing effort does not add token cost.
• Medium: Use when you know what the result should contain and want Claude to produce a complete first version for you to review. Examples include writing a customer follow-up email based on call notes and a price list, turning a video into a post, or preparing meeting minutes and recurring reports. If an answer is generic, first check whether the source material included the necessary information. Raising effort helps only if Claude failed to find or check information that was available.
• High: Use when mistakes may be hard to spot and verification matters, such as checking dashboard totals, validating calculations against source data, reviewing a contract or document sources, and testing reusable instructions (“skills”). The video describes an Anthropic experiment in which higher effort improved success rates in a protein-data analysis. It also suggests a two-stage workflow: draft or make changes with low effort, then verify with high effort.
• Extra high or max: Consider these for complex analyses that require Claude to compare multiple sources, investigate explanations, and work with little user intervention—when the stakes justify the extra time and cost. For example, Claude could analyze why customers cancelled by comparing support conversations and separating customer-stated reasons from its own hypotheses. If both settings are available, the speaker recommends starting with extra high. Max may add cost or time without improving the answer, and further analysis can sometimes overcomplicate a sound solution.

Key Guidance

• Higher effort is not automatically better. It can mean longer waits, more token use, and more decisions made without the user’s input.
• Low effort can be valuable for early drafts because it lets you redirect Claude quickly. In Anthropic’s menu-redesign comparison, low effort produced an interactive draft in about a minute, while max took about 28 minutes.
• Choose effort based on the reasoning required, not simply the amount of material. Sorting many requests into predefined categories may be straightforward, while identifying why customers leave may require deeper comparison—even with fewer conversations.
• If Claude keeps making mistakes despite having the needed information and being asked to check its work, consider switching to a more capable model rather than just increasing effort.
• Anthropic’s general recommendation, as presented in the video, is to start with the model’s default effort and adjust when you need faster responses or more thorough checks.

Sources and Speaker

• Main speaker: Joe, an AI agency owner and educator
• Sources discussed: An article from Anthropic and guidance attributed to Tarik from Anthropic’s Claude Code team