A simple redesign
Quick enough to review and iterate.
The practical AI field guide / 01
Same brain.
Different
effort.
Claude’s effort levels, explained simply. When to turn the dial up, when to turn it down, and why more isn’t always better.
Find the right level ↗Give the task what it needs.
More work can help.
Unnecessary work can wait.
THE IDEA IN ONE LINE
Effort changes how much work Claude puts into the task before coming back to you.
01
Think of a LEGO house.
Same builder. Same instructions. A different amount of time and room to work.
“Make a first version.”
For a simple house, you want something you can see, react to, and change together. Low effort can be a great fit.
“Make sure it all holds up.”
For a complicated castle, the builder may need to test the structure and investigate weak spots. More effort has a purpose.
The model is the builder. Effort is how much work that builder is encouraged to do. Lower effort can reduce capability on difficult tasks; it isn’t just a speed switch.
02
More effort. Same result?
In one simple website redesign, the extra time barely changed the outcome.
A nearly identical result
More time doesn’t guarantee more value.
One anecdotal example from a video walkthrough, not an Anthropic benchmark or a prediction of your results. Bars compare elapsed time in that example only.
Choose the lowest effort that reliably delivers the quality you need.
03
A dial for how you work together.
Use this as a working-style analogy. Move the slider to explore all five levels.
More back-and-forth
Frequent feedbackMore exploration before returning
Room for investigation
Frequent feedbackMore exploration before returning
Room for investigation
Medium
A balanced starting point for clearly defined, everyday work. Give Claude a concrete task and enough context, then review the result.
This line illustrates a collaboration style, not guaranteed behaviour. Effort is not a permission or autonomy setting. Say explicitly when Claude should ask you, and what it may change.
04
Match the effort to the job.
Practical starting points for websites, applications, and integrations. Adjust for your model and the results you observe.
| What you’re doing | Try | Why it fits |
|---|---|---|
| Change headings, colours, or spacing | Low | Small, clear, and easy to review |
| Work together on a website design | Low | Fast feedback helps you steer |
| Build a component with clear requirements | Medium | Several steps with a defined outcome |
| Connect a CMS to an existing site | Medium | Raise it if the integration gets tricky |
| Debug a Cloudflare Worker or API | High | Investigation and edge cases matter |
| Check an order or payment integration | High | A hidden error can be expensive |
| Migrate a large WordPress site to Astro | XHigh | Many dependencies to work through |
| Investigate a particularly difficult problem | Max, selectively | Use when the extra work proves useful |
These are workflow suggestions, not universal model defaults. Available levels and recommended starting points vary by model.
05
Change gears as the project changes.
You don’t need the same effort for the first draft, the final checks, and every spacing tweak.
Build together
LowGet a first version. React to the layout and copy.
Make it work
MediumConnect forms, routes, content, and functionality.
Check the details
HighReview errors, accessibility, and edge cases.
Make small tweaks
LowAdjust headings, colours, and spacing.
Cache behaviour depends on the model and how effort is changed. Supported per-message changes can preserve the API prompt cache; changing the top-level setting can invalidate it. See the official documentation before relying on savings.
06
What does your next task need?
A quick recommendation based on your workflow. Choose the one that best describes your task.
Your starting point
Medium
A balanced place to start for everyday work with clear requirements.
Choose your main priority. This tool is a practical heuristic; your model’s guidance and measured results come first.
07
Put the choice into practice.
In Claude Code, type /effort to open a slider, or add a level to set it directly, like /effort high. Which levels you get, and the default, depend on your model.
/effortClear prompts first.
State the outcome, constraints, and checks you need. More effort won’t fill in every missing requirement.
Low is useful.
Try it for simple changes and quick collaboration. Give multi-part tasks an explicit checklist.
Checking earns its keep.
A task can be small and still deserve careful verification when a mistake would matter.
Want the next guide in your inbox?
THE LATHAM is my weekly newsletter: what I’m building with AI, what’s working, and what the human still has to do.