# A throwaway build can clarify a feature before planning Markdown copy of https://realalpha.app/s/a-throwaway-build-can-clarify-a-feature-before-planning/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy. - Speaker: [Dan Shipper](https://realalpha.app/people/dan-shipper/) (Host) · Co-founder & CEO, Every · Sold his first startup, Firefly, to Pegasystems; writes Every's Chain of Thought. - Show: AI & I - Episode: [The Secrets of Claude Code From the Engineers Who Built It](https://realalpha.app/v/IDSAMqip6ms/), Oct 29, 2025 - Moment: 15:46 to 20:29 of the episode (5 min), playing on https://realalpha.app/s/a-throwaway-build-can-clarify-a-feature-before-planning/; the episode from there: https://realalpha.app/v/IDSAMqip6ms/?t=946 ## Quote > “I have Claude just like implement it just **without giving it anything else**.” > — [Dan Shipper](https://realalpha.app/people/dan-shipper/) at [16:17](https://realalpha.app/v/IDSAMqip6ms/?t=977) · AI & I ## Takeaways 1. Try a first implementation before committing to a full specification. 2. Use the mistakes or surprises to sharpen the real plan. 3. Switch to plan mode for harder, multi-hour feature work. ## About this segment Dan Shipper uses Claude to make a first pass from a vague request, then turns its mistakes into a better spec. Cat and Boris describe when to prototype and when to plan. From “The Secrets of Claude Code From the Engineers Who Built It”, Oct 29, 2025. [Watch the full episode](https://realalpha.app/v/IDSAMqip6ms/) ## Transcript From the transcript AI & I publishes with the episode. [15:51](https://realalpha.app/v/IDSAMqip6ms/?t=951) I would love to go deeper into the... How do you make a good plan? So the sort of the feature dev thing, because I think there's a lot of like little tricks that I'm starting to find or people at every start starting to find that work. And I'm curious, like what, what are the things that we're missing? So for example, one step in the, one unintuitive step of the, you know, plan development process is, Even if I don't exactly know what the thing that needs to be built is, I just have like a little sentence in my mind, like I want feature X. I have Claude just like implement it just without giving it anything else. [16:26](https://realalpha.app/v/IDSAMqip6ms/?t=986) And I see what it does. And that helps me understand like... okay, here's actually what I mean because it made all these different mistakes or it did something that I didn't expect that might be good. And then I use that, like the learning from the sort of throwaway development. I just clear it out, and then that helps me write a better plan spec for the actual feature development, which is something that you would never do before because it would be too expensive to just YOLO, send an engineer on a feature that you hadn't actually specced out. [16:53](https://realalpha.app/v/IDSAMqip6ms/?t=1013) But because you have Cloud going through your code base and doing stuff, you can like learn stuff from it that helps inform the actual plan that you make. Yeah, I feel maybe I can start and I'm curious how you use it too. I think there's like a few different modes maybe for me, like one is prototyping mode. So like traditional engineering prototyping, you want to kind of build the simplest possible thing that touches all the systems just so you can kind of get a vague sense of like what are the systems? [17:19](https://realalpha.app/v/IDSAMqip6ms/?t=1039) There's unknowns and just to kind of trace through everything. Yeah. And so I do the exact same thing as you, Dan. Like, Claude just does the thing, and then I see where it messes up, and then I'll ask it to just throw it away and do it again. So just hit escape twice, go back to the old checkpoint, and then try again. I think there's also maybe two other kinds of tasks. So one is just things that Claude can one-shot. [17:41](https://realalpha.app/v/IDSAMqip6ms/?t=1061) And I feel pretty confident it can do it. So I'll just tell it, and then I'll just go to a different tab, and I'll shift tab to auto-accept, and then just go do something else or go to another one of my Claude's and tend to that while it does this. Yeah. But also there's this kind of like harder feature development. So these are, you know, things are maybe in the past, it would have taken like a few hours of engineering time. [18:01](https://realalpha.app/v/IDSAMqip6ms/?t=1081) And for this usual, I'll shift tab into plan mode and then align on the plan first before it even writes any code. And I think what's really hard about this is the boundary changes with every model in kind of a surprising way. where the newer models, they're more intelligent. So the boundary of what you need plan mode for got pushed out like a little bit. Like before you used to need to plan, now you don't. [18:23](https://realalpha.app/v/IDSAMqip6ms/?t=1103) And I think this is general trend of like stuff that used to be scaffolding with a more advanced model, it gets pushed into the model itself. And the model kind of tends to subsume everything over time. Yeah. How do you think about like building a agent harness that isn't just going to like you're not spending a bunch of time building stuff that is just going to be subsumed into the model in three months when the new cloud comes out? [18:46](https://realalpha.app/v/IDSAMqip6ms/?t=1126) Like, yeah. How do you how do you know what to build versus what you just say? It doesn't work quite yet, but next time it's going to work. So we're not going to spend time on it. I think we build most things that we think would improve quad code's capabilities, even if that means we'll have to get rid of it in three months. If anything, we hope that we will get rid of it in three months. I think for now, we just want to offer the most premium experience possible. [19:12](https://realalpha.app/v/IDSAMqip6ms/?t=1152) And so we're not too worried about throwaway work. Interesting. Yeah, and an example of this is something like even like plan mode itself. I think we'll probably unship it at some point when Claude can just figure out from your intent that you probably want to plan first. Or, you know, for example, I just deleted like 2000 tokens or something from the system prompt yesterday. It's because like Sonnet 4.5 doesn't need it anymore. [19:35](https://realalpha.app/v/IDSAMqip6ms/?t=1175) But Opus 4.1 did need it. What about, you know, in the case where the latest Frontier, you know, model doesn't need it, but, you know, you're trying to figure out how to make it more efficient because you have so many users that, you know, you're maybe you're not going to use Opus or Sonnet 4.5 for everything. Maybe you're going to use Haiku. So there's a tradeoff between having a more elaborate harness for Haiku versus just like not spending time on it, using Sonnet, eating the cost and working on more Frontier type stuff. [20:05](https://realalpha.app/v/IDSAMqip6ms/?t=1205) In general, we've positioned QuadCode to be a very premium offering. So our North Star is making sure that it works incredibly well with the absolutely most powerful model we have, which is Sonic 4.5 right now. We are investigating how to make it work really well for future generations of smaller models. But it's not the top priority for us.