Blog
rewrite codebase AIstartup code rewriteOpus 4.5 rewrite

Should You Rewrite Your Codebase With AI? Octar's Lessons

A startup built for every sales demo and ended with a product nobody could learn. Here is when an AI-assisted rewrite makes sense, and when it doesn't.

Portrait of Luis Chavez-Mattos
Edited by Luis Chavez-Mattos, Director, Product··RSS

For Engineering, Product Management and Executives

A tiny man seen from directly above pushes one huge stack of blank pages onto an almost identical stack on an empty flat expanse.

Should you rewrite your codebase with AI?

Sometimes, but only if the problem is the shape of the product, not just messy code. AI coding models have made rewrites faster and cheaper than they used to be, which changes the math. It does not change the reason to rewrite. A rewrite makes sense when the product's structure fights your users, when patching it costs more than rebuilding it, and when you know what the new version should be. Octar, a Y Combinator startup, is a useful example of all three.

TL;DR

  • A rewrite fixes structure, not taste. If the real problem is a confusing product, new code alone will not fix it.
  • Building a feature after every sales call creates a monster. Octar's founders say their product became a "Frankenstein of demo features" that customers struggled to adopt.
  • Cheaper rewrites change the decision, not the discipline. Stronger AI coding models lower the cost of starting over, but you still need a clear design for the new version.
  • Many one-off features usually hide a few general ones. Octar had separate builders for two kinds of documents when one flexible document builder would have done the job.
  • Protect your customers during the switch. Plan what stays the same, what changes, and how users get from old to new.
  • Treat reported results as one company's story. Octar's own account describes a rebuild in about two months and several times more usage, which is encouraging but not a guarantee.

What is Octar and what went wrong?

Octar sells an AI platform for large enterprise software implementations, the kind of project where a big company rolls out a system like ServiceNow, Salesforce or SAP. The founders say these projects often fail or run over budget because of misalignment in requirements and design at the very start, before anyone configures anything. So Octar started upstream, helping teams keep requirements, design and configuration connected from first discovery through go-live.

The founders did real homework. One of them spent about two months talking to dozens of services leaders, system integrators and internal IT teams. Around the 30th conversation, he wrote down what he thought the AI models could solve. His first idea was configuration agents for each platform. When he showed it to roughly 50 people, they said the real pain was upstream, in requirements and design. That reframing became the company.

Then came Y Combinator (spring 2025 batch). The founders launched early, including a LinkedIn post featuring a golden retriever that they say reached close to a million impressions and drove a few hundred demo bookings in three days. Many of those were not qualified buyers, but some large enterprise contracts came from that attention.

The trouble started after the demos.

Why did building for every demo backfire?

The founders describe a very tight loop. One of them took sales calls, and the two builders sat across from him. They heard the pushback live, and that night they shipped something to address it. By their own description, they built for the demos "over and over and over again."

It worked for selling. It did not work for using. In their words, they were reacting on gut instinct and not digging into why a prospect asked for something. After a call, they might ship three new features, and the prospect would say it "nails it." But when customers actually came onto the platform, the learning curve was too high.

The clearest example they gave: if a customer wanted a business requirements document and a solution design document, Octar had two separate features, a BRD builder and a solution design builder. What the customers needed was one document builder with a good agent harness (the surrounding logic that lets an AI agent do the work) that could adapt to whatever output they wanted.

This is a common pattern, and it is worth recognizing in your own work:

  1. A customer asks for a specific thing.
  2. You build exactly that thing.
  3. The next customer asks for a slightly different thing.
  4. You build that too, as a separate feature.
  5. After 50 rounds, nobody can find anything, and each feature is a small special case.

The problem is not listening to customers. It is treating each request as the answer instead of as evidence of an underlying need.

What does a full rewrite look like with AI models?

The part of the story I can describe with confidence is the diagnosis. For the rewrite itself, I am relying on the company's own account, which ties the decision to the release of Anthropic's Opus 4.5 model. That account describes rebuilding the whole product in roughly two months and then seeing usage rise five to ten times. Those are the company's figures and I have not checked them independently. Take them as one data point, not a benchmark.

What the diagnosis tells you about the shape of a good rewrite is more useful than the numbers:

  • Fewer, more general features. Replace a pile of single-purpose tools with a smaller set that can flex. One document builder instead of many.
  • A design you can state in a few sentences. If you cannot explain what the new product is without listing features, you are not ready.
  • Customer knowledge carried over. Octar spent its early weeks learning the problem deeply. The rewrite keeps that knowledge. It throws away the accumulated one-offs.

AI models help because the slow, expensive part of a rewrite used to be typing out the new code. Models that handle large chunks of work shrink that part. They do not shrink the thinking part: deciding what to build, what to cut and what to keep.

When does a rewrite make sense?

Use this as a checklist. A rewrite is worth considering when most of these are true.

QuestionPoints toward a rewritePoints toward fixing in place
Why do users struggle?The structure is confusing (too many overlapping features)A few specific screens or bugs
How does new work feel?Every change touches many unrelated partsChanges stay local
Do you know the target?You can describe the simpler product clearlyYou are still guessing what it should be
Who depends on it?You can migrate customers in stagesCustomers rely on exact current behavior
Cost of starting overFalling, because AI tools do much of the buildStill high, or the code holds hard-won edge cases

If you are in the right-hand column on most rows, patch and refactor instead. Rewrites have a long reputation for going badly because old code often contains quiet fixes for problems no one remembers. A rewrite can lose those.

What are the risks of rewriting everything?

  • Lost edge cases. Old code encodes years of small fixes. Write down the behaviors that matter before you start, and test against them.
  • Two products at once. While you rebuild, customers still use the old one. Decide how you will support both and for how long.
  • Scope creep in the new version. The same instinct that created the mess (say yes to everything) can recreate it. Set a short list of what the new version does.
  • Speed that hides mistakes. AI-generated code arrives quickly, which makes it easy to skip review. Keep tests and human review in the loop.
  • Customer trust. Users who just learned the old product need a clear explanation of what changes and why.

How do you keep customers during a rewrite?

The material I have does not spell out Octar's exact customer plan, so here is general practice you can apply:

  1. Talk to your most engaged customers first. Ask what they actually use, not what they asked for. Usage data and short calls tell you which features to fold together.
  2. Keep the outcomes, change the path. Customers care that their requirements document gets produced, not that it came from a specific button.
  3. Migrate in stages. Move a small group first, fix what breaks, then widen.
  4. Give a clear date and a clear reason. "We made it simpler so it is faster to learn" is a reason customers understand.
  5. Keep data intact. Whatever customers created in the old version must show up in the new one.

What does this mean for you?

If you run or work on a product team: Look at your feature list. Find the features that are near-duplicates and ask whether one flexible feature could replace them. That is often a smaller project than a full rewrite and captures much of the benefit.

If you are a founder or manager under sales pressure: Before building a requested feature, ask what problem the customer is trying to solve and whether a feature you already have could cover it. Keep a running list of requests grouped by underlying problem, not by wording.

If you are not technical: You may still be the person deciding whether a rewrite is worth the money. Ask your team three things: what specifically is wrong with the current structure, what the new version will not do, and how customers will move over. Vague answers are a warning.

If you depend on software you bought: Vendors with cluttered products are often feeling the same pressure. Ask them about their roadmap for simplification, and notice whether new features are replacing old ones or only adding to them.

Can you build your own feature request tracker?

If the lesson is to group requests by the problem behind them, you need somewhere to put them. Instead of a spreadsheet nobody maintains, or another subscription, you could describe a small internal tool to Remy in plain English: a form where sales and support log what a customer asked for, a field for the underlying problem, a tag for the existing feature that might already cover it, and a dashboard showing which problems come up most. Remy builds it as a full-stack app with sign-in and a database, and the code is yours to keep. It will not decide what to build, but it makes the pattern visible.

Frequently Asked Questions

4 questions

Is a full rewrite ever a good idea?

Yes, when the product's structure is the obstacle and you have a clear picture of the simpler version. It is a poor idea when the problems are specific bugs or when you are not sure what the new version should do. Cheaper AI-assisted coding lowers the cost but does not remove the risk.

What does "demo-driven" development mean?

It means building features mainly to win the next sales call. Each request gets its own feature, quickly. The result sells well in a demo but can be hard for real users to learn, which is what Octar's founders say happened to them.

Did AI make Octar's rewrite possible?

By the company's own account, the rewrite followed the release of Opus 4.5 and took about two months. The reported results, including several times more usage, are the company's claims. Stronger models likely made starting over more practical, but the key decisions about what to build were still human ones.

How do I know if my product has too many features?

Warning signs include near-duplicate features, long onboarding for new customers, and users who only touch a small part of the product. If customers say they love a feature but never use it after the demo, that is a strong signal.

How we source, write, and correct our articles. Editorial standards

Related articles

Straight to your inbox.
New pieces as they publish. No noise.