When people start working with Claude, they quickly discover the effort setting next to the model picker. Almost everyone reacts the same way: they push it up to High or Max. It seems logical, since important work deserves the maximum.
However, this setting doesn't improve the quality of the answer; instead, it changes how much work Claude puts into a response. Whether this pays off depends on the task. In this article, I summarize what effort actually does, what each level is for, and how our custom-developed mod can take the decision off your shoulders.
What does effort skill actually do?
Effort doesn't make the model smarter. The same knowledge base is behind it at both Low and Max levels. The difference lies in how much Claude reads before starting, how many times it checks its own work, how many edge cases it tests, and how much it decides on its own instead of asking for clarification.
This comes at a price: more verification takes more time and costs more tokens. For the same request, the high level can use up to seven times as many tokens as the low level. (bold: "seven times as many tokens")
An example
Suppose the monthly revenue in a report is 0.4% lower than in the accounting records.
Low effort
Claude reviews the query and finds an obvious cause, for example that cancelled invoices are not filtered out. It fixes it, and most of the gap disappears.
High effort
It does not stop at the first explanation. It compares the two sources day by day, checks for duplicate rows in the load, and looks separately at the last day of the month.
That is how the cause of the remaining gap comes to light: the system closes the day in UTC, so invoices issued on the last evening of the month slip into the next month. The low level gives a fast and convincing first answer, but the high effort level also finds the second, silent error.
This leads to two rules of thumb:
01
Save high effort for places where a hidden error can be costly.If you review the result yourself anyway, a higher level only makes things slower and more expensive.
02
More effort does not help if the direction is wrong.If Claude does not understand the task, it needs context. If the task exceeds its capabilities, it needs a stronger model.
The question is always whether it did not know enough or did not try hard enough.
Which effort level is for what?
LowYou are part of the loop, or the task is simple
Claude quickly gives a usable answer, and you decide and iterate. Here speed is worth more than thoroughness.
- ›Explain what a MERGE statement does in SQL.
- ›Give me five ideas for the presales workshop agenda.
- ›Rephrase this sentence to sound friendlier: “We can’t meet the deadline unless you resolve the dependencies.”
- ›Rename the
cust_id column to customer_id.
MediumEveryday work
Claude works from clear input: an outline, notes, a specification. If the input is good, a higher level rarely adds much.
- ›Create a 10-slide presentation on the Q3 results based on the attached outline.
- ›Write a specification of the customer portal login flow based on the meeting notes.
- ›Build a Power BI report on daily sales based on the given specification.
- ›Briefly look up what Microsoft Fabric Direct Lake mode can do.
HighWhere a hidden error is expensive
Production systems, numbers, money, contracts, permissions, migrations. This is where it really pays off that Claude checks, looks for edge cases and investigates discrepancies before delivering a result.
- ›Fix the bug in the existing invoice export: in some rows the total is off by a few cents.
- ›Review the SLA clauses of the draft contract from a risk perspective.
- ›Check the row-level security settings in the Power BI dashboard.
- ›Migrate the SCD2 dimensions to the new Databricks lakehouse.
Extra (xhigh)When you hand it off
Long, complex work that Claude has to carry through and verify on its own, because nobody will be around to answer its questions.
- ›Build the full ingestion pipeline from the three source systems on your own, test it end to end, and don’t ask questions along the way.
- ›Run the full regression test overnight, fix the bugs, and have the report ready by morning.
Max is the most extreme version of this: it is meant for rare, critical tasks where every additional check matters. In everyday work it is almost never needed, which is why the mod I built never chooses it on its own.
UC Effort Optimizer: leave the decision to it
In practice, no one manually switches effort levels before every prompt. That’s why I created the UC Effort Optimizer. It is a Claude Code mod that anyone can download and install, and it makes this decision for you with every prompt.
Claude Code
YouFix the bug in the existing invoice export: in some rows the total is off by a few cents.
1Effort switched to High (was Medium) · production financial data, possible hidden rounding error, verification needed
> Type your message…
AutoOpus · Effort High 2
1 When it switches, a line above the prompt shows for 15 seconds what the mod changed to and why. 2 The actual level is always shown next to the model name, color-coded.
What does it do?
It reads every prompt before Claude starts.It runs a short assessment based on Anthropic’s effort recommendations and picks Low, Medium, High or Extra. It starts from the model’s default level and only deviates from it when there is a reason.
It tells the topic apart from the task.“Explain the security model!” stays Low, while “Check the security settings!” becomes High. It does not confuse missing knowledge with too little effort: an architecture decision, for example, calls for a stronger model, not more checking.
It learns from your feedback.If you say Claude skipped files or didn’t run the tests, it moves up one level. If the answer was too long or too slow, it moves down one level.
The chosen level sticks.It stays in effect for the following prompts until a new assessment picks something else. Replies like “ok, go ahead” or “option 2” don’t move it.
You always see what Claude is running on.The actual level is shown next to the model name, color-coded: Low grey, Medium green, High orange, Extra red. When the mod switches, a line above the prompt shows for 15 seconds what it switched to and why.
Switch modes with one click.In the desktop app, click the level indicator to choose from three modes. In Auto mode it decides on its own. In Ask mode it asks before switching. In Manual mode it doesn’t touch anything.
UC Effort Optimizer
Current level: Medium Chosen by Auto
AutoAskManual
Auto mode: the mod picks the right effort level for every prompt on its own and tells you when it switches.
LowMediumHighExtra
Clicking the level indicator opens the mod’s panel: it shows the current level, and this is where you switch modes.
Why is it good?
You don’t have to think about it.Good practice applies automatically, and consistently across the whole team. It doesn’t depend on who is paying attention.
It spends where it matters.Time and tokens go into the checks that find real errors, not into overthinking a simple explanation.
Transparent.You always see what level Claude is running at, and why it switched.
You stay in control.If you change the effort picker by hand, the mod steps back and your choice applies until you switch Auto mode back on.
It keeps costs in check.It never picks Max on its own and never touches your saved settings.
How much does it cost?
Almost nothing. It runs a quick check per prompt, using roughly a thousand tokens and taking about a second. Unlike previous models, switching doesn't cost extra: current Claude models don't re-read the entire context when changing effort levels, as the cached part of the conversation remains. We measured it on a 5-hour task:
With the built-in statistics (/uc-effort-optimizer stats), anyone can check it on their own work. (code format: "/uc-effort-optimizer stats")
Summary
Effort is not a quality slider, but a budget for verification. The most expensive errors are often hidden. A duplicate active record in an SCD2 dimension, a late-arriving item, or an exchange rate falling on the wrong day all silently distort results without triggering an error message. You need the high level for tasks like these.
“Everywhere else, a good brief and fast iteration are cheaper, and often give better results.”
If you don't want to decide for yourself which effort level to choose for every prompt, leave it to the UC Effort Optimizer.
Download: UC Effort Optimizer
A Claude Code mod under the MIT license. Drag the downloaded package into a Claude Code chat and ask Claude: “Install this mod.” Claude will first ask where to put the mod. Choose a permanent location, because the folder must not be deleted after installation. From then on, the mod loads automatically every time Claude starts. Detailed steps are in the README included in the package.
