Darius: You ask an AI to write a piece of security code. A filter to strip bad stuff out of a web page. On the 'low effort' setting, it takes two minutes, writes the code, and calls it a day.
Tessa: Okay, standard.
Darius: You ask it again, but this time you flick the dial to 'high effort'. It takes thirty-three minutes. And in that time, it writes the code... then it adversarially reviews its own draft. Then it reads the source code of the underlying library it's using to check for bugs in that. And then it writes a fuzzer to randomly generate attacks against its own code to see if it can break it.
Tessa: Wait. It attacks itself?
Darius: It becomes its own red team. And this is all from a blog post from the Claude team about this feature in their coding assistant called 'effort'.
Tessa: So 'high effort' just means 'be paranoid'.
Darius: Essentially, yeah. The author describes it as a way to tell the model how much compute you want it to spend. Like, if you ask a person to do a task in an hour, they'll give you a decent first pass. If you give them twelve hours, they'll assume you want them to be incredibly thorough, check all the corner cases.
Tessa: But isn't that what I want the AI to do all the time? Be thorough? Why do I have to ask it to try hard? That feels like the default setting should be... good.
Darius: Because you don't always need that thirty-three-minute paranoid version. The post gives an example of building a fitness app. On low effort, it builds a simple log and a graph. A starting point. On max effort, it adds heat charts and all this other complexity you might not have wanted. It makes more assumptions on your behalf.
Tessa: So it's a trade-off between speed and the AI going rogue with its own ideas.
Darius: Exactly. The author suggests a workflow: get the AI to interview you about the feature, then have it implement a first draft on low effort. You review it, iterate quickly, and then—only when you think you're close—you crank it up to high effort to verify and test for all the edge cases you didn't think of.
Tessa: Okay, that makes some sense. You do the creative part fast and loose, then you bring in the robot for the final inspection.
Darius: And the inspection is where it gets wild. They tested this on a benchmark called Terminal-Bench, which has these insane tasks. Like, 'build an 8-bit game console in Verilog'. Or 'formally prove Takens’ embedding theorem in Lean 4'.
Tessa: These are not my usual software engineering tasks.
Darius: Right. And for these hard tasks, effort makes the difference between a pass and a fail. There's one where it has to fix a bug in a storage engine. At low effort, it just edits the code without even running the tests. At high effort, it first reproduces the crash, then writes its own randomized tests to make sure the fix is robust.
Tessa: I'm sorry, I'm stuck on this. It feels like a bug they turned into a feature. Like, the AI's default state is to be sloppy and cut corners, so they give us a dial to bribe it into doing its job properly.
Darius: I see it differently. I don't think it's about being sloppy, it's about being appropriate. You don't want the AI spending an hour writing a fuzzer when you just asked it for a quick sketch of a user interface.
Tessa: But how am I supposed to know when to use it? I'm not a security expert. I don't know my problem has a million hidden edge cases until the AI fails to solve it. I'm just going to set it to 'Max' every time, get annoyed that it's slow, and then get even more annoyed when it's still wrong because the article itself says high effort can't fix a fundamentally bad approach.
Darius: But maybe that's the new skill. The new job is not just writing the prompt, it's diagnosing the type of problem you have. Is this a wide-open creative task, or a narrow, deep, security-critical one? You're not just a user anymore, you're a manager of this weird, non-human intelligence.
Tessa: A manager of a very weird, very literal employee. It sounds exhausting. I'm not sure I'm buying that this is a step forward and not just... more work.
Darius: It might be more work. Or just different work. The piece ends with a simple rule of thumb: Low effort for brainstorming and easy stuff. Medium for normal features. High for bug fixes in old codebases. And Max for when you want it to go full autonomous-agent mode and solve a hard problem from end to end.
Tessa: Hmm. A dial for paranoia. Okay, I'm still sitting with that.
Darius: Me too. I'm Darius.
Tessa: I'm Tessa. This has been the audiclip podcast.