復活したClaude Fable 5に、仕様書なきレガシーシステムを「発掘」させてみたAn experiment using Claude Fable 5 to reverse-engineer an undocumented legacy…

Legacy systems that run critical business processes but lack any written documentation are one of the most persistent problems in enterprise software. A recent experiment, published on Zenn, put a model referred to as the revived Claude Fable 5 to work on exactly this challenge, tasking it with analyzing the source code of an undocumented legacy system and reconstructing a specification from scratch. The reported outcome suggests that the model can "excavate" buried business logic effectively enough to offer a practical starting point for modernization efforts.

The core problem the experiment addresses is familiar to anyone maintaining aging codebases. Over years or decades, systems accumulate rules, edge cases, and workarounds that exist only in the code itself. The original authors move on, comments go stale or were never written, and the institutional knowledge of why a particular calculation or branch exists simply evaporates. When a company later wants to migrate, refactor, or replace such a system, the first and often hardest step is figuring out what the software actually does. This is the gap the author set out to test whether an AI model could help close.

According to the write-up, the approach was essentially a form of AI-assisted reverse engineering. Rather than asking the model to rewrite or fix the code, the experiment focused on comprehension: feeding it source code and prompting it to identify the underlying business logic, describe the data flows, and produce human-readable specifications. The emphasis on "excavation" is notable. In a legacy system, the meaningful rules are frequently tangled together with dead code, defensive checks, and implementation details that no longer matter. Distinguishing the intended behavior from incidental clutter is the part that traditionally requires an experienced engineer, and it is where the model's language-and-code reasoning appears to add value.

The claimed result is that Claude Fable 5 was able to surface and reconstruct logic that had been effectively invisible, turning opaque code into a described specification. The author frames this as a demonstration rather than a finished methodology, which is an appropriate level of caution. Reconstructing intent from implementation is inherently interpretive, and an AI-generated specification is a hypothesis about what the code means, not an authoritative record. Any output of this kind would still need verification against the running system, test cases, or surviving domain experts before it could be trusted for a migration.

It helps to place this in the broader industry context. Using large language models to read and explain existing code has become one of the more widely adopted uses of AI in software development, alongside code generation and completion. Tools such as GitHub Copilot, Cursor, and various IDE integrations already offer "explain this code" features, and Anthropic's own Claude models have been positioned around large context windows that make it feasible to ingest substantial amounts of source at once. For legacy work specifically, the ability to hold and reason over a large body of interrelated code is arguably more important than raw generation quality, because understanding a legacy system requires seeing how distant parts connect.

Reverse engineering and legacy modernization also sit within a long tradition of tooling that predates generative AI. Static analysis tools, dependency mappers, and documentation generators have existed for years, and structured approaches like the strangler fig pattern describe how to incrementally replace an old system while it keeps running. What an approach like the one described here seems to add is a way to bridge from raw code to natural-language understanding faster than manual reading allows, potentially lowering the cost of the initial discovery phase that these established methods depend on.

Several caveats are worth keeping in mind. Feeding proprietary source code to a hosted model raises data governance and confidentiality questions that any organization would need to resolve before attempting this at scale. There is also the risk of plausible-sounding but incorrect summaries, since a model may confidently describe logic that does not match actual behavior. As a single experiment on one system, the results are suggestive rather than conclusive. Still, the exercise illustrates a genuinely useful direction: rather than replacing the difficult human judgment involved in modernizing old software, AI models like this appear capable of accelerating the most tedious part of it, giving engineers a documented map of territory that was previously unexplored.