---
title: "Claude Code Modernization Plugin: How It Proves a Rewrite Works"
date: "2026-11-14"
video_id: "7708091083342826785"
tiktok_url: "https://www.tiktok.com/@therobertta_/video/7708091083342826785"
creator: "@therobertta_"
category: "Coding"
tags: ["claude code", "code modernization", "claude plugins", "legacy code", "anthropic"]
excerpt: "Anthropic's code-modernization plugin rewrites legacy code and proves it behaves the same with byte comparisons, invented inputs and a canary. How it works."
---
The code-modernization plugin in Anthropic's claude-plugins-official repository points Claude at a legacy codebase and produces an understanding of it, a plan you approve, the rewritten code, and script-checked proof that the new code behaves like the old. A TikTok from Robert Ta (@therobertta_) reads its README accurately. The plugin was not new when he posted it on 14 November 2026, and the video skips a few limits on what "proof" means.

## What is the Claude code-modernization plugin?

The plugin lives at plugins/code-modernization in the anthropics/claude-plugins-official repository on GitHub. Its README opens with "Point Claude at a legacy codebase and get four things":

1. **An understanding:** an assessment, a dependency map, and business rules written as Given/When/Then cards with file:line citations.
2. **A plan** that you must approve before anything is built.
3. **The modernized code.**
4. **Proof** that the new code behaves like the old, produced by a step called modernize-verify.

The README says it handles any language and any kind of move: a newer version, a rewrite in another language, or a rebuild on a new architecture. The video does not mention that scope.

## Is the plugin new?

The video says "Claude Plugins Official in their repo just dropped an official code modernization plugin." The GitHub page on screen shows the last commit, by morganl-ant, as "3 weeks ago."

The commit history for the folder lists 35 commits under the date Sep 25, 2026, all by morganl-ant. The newest titles are "count failures and the machine's python, once per version" and "version 1.0.0, usage counts and a README results table." Older commits sit on further pages, so the first commit date is unknown. The plugin had been in the repository and reached version 1.0.0 before the video was posted.

## How does it prove the rewrite works?

The video says the proof is "built from evidence a script can check, not from the model's opinion." The README describes several controls.

**Byte comparison.** Where the legacy code can run, old and new code run on the same inputs, and a script compares every byte of output. Fields that must vary, such as timestamps and generated IDs, are masked, and each masked field is listed. A commit title notes that floating-point output is compared within a declared tolerance. So "every byte" holds after the listed exceptions.

**Invented inputs.** The verifier invents at least ten new inputs and compares again. They cover boundaries, empty and huge inputs, unusual ordering, and malformed records when exact behavior is requested.

**The canary.** A deliberate one-line break must turn the tests red. This shows the test suite is capable of failing.

**Checking the checker.** The comparison is tested by changing one byte of an output to confirm the script catches it.

**Clean re-runs.** Tests are re-run from clean. A run that executed no tests, or skipped them, counts as a failure.

**Critical rules.** Each critical (P0) business rule must be named by a test that ran and passed. Rules backed only by skipped tests are reported as "named, not run." The video says "every critical rule is backed by a test that ran," which matches.

## What verdicts does it give?

Each module gets one verdict from scripts/proof_pack.py: PROVEN, PARTLY PROVEN or NOT PROVEN. If the legacy system cannot run in your environment, the best possible verdict is PARTLY PROVEN, because there is nothing to compare against byte for byte. The video does not mention that ceiling.

## What the canary does and does not show

The canary proves the tests can fail. It does not prove the tests cover what the legacy code actually does. The invented-input comparison against the running legacy system is the control aimed at that second question, which is why a legacy system that cannot run caps the verdict.

## What the video does not show

The speaker reads the README aloud. There is no run of the plugin on a real codebase, no verdict output and no failure rate. He closes with "seems to be very well thought out from what I'm reading here," which is a reading, not a test. The README has a results table, according to a commit title, but the video does not cite it.

The description says this is how "most legacy rewrites go wrong without anyone noticing." Neither the video nor the README gives a source for "most."

The description links a longer YouTube video, which was not reviewed here.

## Bottom line

The plugin's README matches the video: four deliverables, byte comparison of old and new output, at least ten invented inputs, a canary and a one-byte check of the comparison. The byte comparison excludes listed masked fields, and a legacy system that cannot run caps the result at PARTLY PROVEN. The plugin had 35 commits and a 1.0.0 version weeks before the video called it new.

## Sources

- [code-modernization README, anthropics/claude-plugins-official](https://github.com/anthropics/claude-plugins-official/blob/main/plugins/code-modernization/README.md): the four deliverables, verification controls, canary and verdicts.
- [Commit history for plugins/code-modernization](https://github.com/anthropics/claude-plugins-official/commits/main/plugins/code-modernization): commit titles, the Sep 25, 2026 date, the author and version 1.0.0.
- [Plugins directory, anthropics/claude-plugins-official](https://github.com/anthropics/claude-plugins-official/tree/main/plugins): the code-modernization folder.


## Video description and verification notes

**The clip runs 68.08 s of video and 68.10 s of audio. It is 576x1024 (9:16) H.264 High at 24 fps, about 638.5 kbps. The audio is HE-AACv2 stereo at 44.1 kHz, about 32 kbps. The container is MP4 (isom), the overall bitrate is 675,531 bps and the file is 5,750,126 bytes. metadata.json gives format id h264_540p_675531-1, but its fps field is null, so the 24 fps figure comes from ffprobe. The post timestamp is 1794679805, which is 2026-11-14 UTC, and the metadata capture was 2026-11-14T21:01:22Z. The uploader is @therobertta_ (nickname Robert Ta), and the audio is credited as "original sound" by Robert Ta. The description is 391 characters long and links a YouTube full video. Counts at capture were 1,362 views, 66 likes, 1 comment, 16 reposts and 66 saves. I read one frame, frame_001 (about 0 to 2 s), and no others, so my visual evidence is limited to that opening frame. It shows a GitHub page for claude-plugins-official, path plugins/code-modernization/README.md, with the heading "Code Modernization". The commit row reads "morganl-ant", with the commit hash "c8874f5" and the age "3 weeks ago". The README is listed as 220 lines (162 loc) and 30.2 KB. Below the page is the burned-in caption "Anthropic dropped a Code Modernization plugin". Under that sits a partial live caption, "teams,". At the bottom are two webcam tiles: a long-haired man on the left and a man with earbuds on the right. I did not read the later frames, so I cannot describe the rest of the screen content, and I make no claim that a terminal or a run of the plugin appears. The transcript is 24 lines of text and 23 Whisper segments (ids 0 to 22), 178 words by my rough count. The speaker reads the README aloud and finishes by calling the verification design "very well thought out". Segment 22 ("here.") runs 67.70 to 68.06 s, inside the 68.10 s audio, so the transcript does not run past the end of the audio. It is a short trailing fragment, and its decode used temperature 0.8, which is the only sign of Whisper uncertainty.**

## Claim: "dropped" an official plugin (partly corrected on timing)

At 4.50 s the video says "Claude Plugins Official in their repo just dropped an official code modernization plugin." The repo is real. The plugin folder exists at plugins/code-modernization in the anthropics/claude-plugins-official repository. The on-screen commit age of "3 weeks ago" in frame 1 already contradicts "just". The GitHub commit history page for that folder (fetched through a summarizer) lists 35 commits under a single date header, Sep 25, 2026, all by morganl-ant. The newest titles are "count failures and the machine's python, once per version" and "version 1.0.0, usage counts and a README results table". The fetched view showed a "Next" link to older commits, so the first commit date is unknown to me. The plugin was therefore in the repo before the video was posted on 14 November 2026. The commit list shows 35 commits and a 1.0.0 version, which suggests iteration before the video, not a first drop. The summarizer gave no per-commit timestamps, so I treat the Sep 25 date as indicative and not exact.

## Claim: four deliverables (verified)

At 13.24 s and 17.54 s the video lists an understanding, a plan you approve, the modernized code, and proof that the new code behaves like the old. The README matches this. It describes an understanding (assessment, dependency map, and business rules written as Given/When/Then cards with file:line citations). It also describes a plan that must be approved before anything is built, the modernized code, and proof produced by a step called modernize-verify. Frame 1 shows the README sentence "Point Claude at a legacy codebase and get four things" with the same list. The frame also adds that the plugin handles any language and any kind of move: a newer version, a rewrite in another language, or a rebuild on a new architecture. The video does not mention that scope.

## Claim: proof is script-checked, with byte comparison and 10 invented inputs (verified)

At 29.36 s the video says the proof is "built from evidence a script can check, not from the model's opinion". It then says old and new code run on the same inputs and "a script compares every byte" (39.82 s), and that the verifier "invents at least 10 new inputs" (42.16 s). The README says the same: where the legacy code can run, a script compares every byte of output. It masks only fields that must vary, such as timestamps and generated IDs, and lists each masked field. It also says the verifier invents at least ten new inputs, covering boundaries, empty, huge, unusual order, and malformed records when exact behavior is requested, and then compares again. The video omits the masking exception. The omission matters, because "every byte" is true only after the listed masked fields are excluded. The plugin also compares floating-point output within a declared tolerance, according to a commit title, so the video's byte-for-byte phrasing is a simplification for numeric output.

## Claim: the canary and the one-byte check (verified, with added context)

The description says the canary "breaks one line on purpose to make sure your tests actually turn red". At 54.0 s the video says "A deliberate one-line break must turn tests red, the canary", and at 58.78 s that the comparison is checked by "changing one byte of an output". The README says the same on both points. Two further README controls get no mention in the video. Tests are re-run from clean, and a run that executed no tests or skipped them counts as a failure. Each critical (P0) rule must be named by a test that ran and passed, and rules backed only by skipped tests are reported as "named, not run". The video does say at 63.22 s that "every critical rule is backed by a test that ran", which matches the second control. Each module then gets one verdict from scripts/proof_pack.py: PROVEN, PARTLY PROVEN or NOT PROVEN. If the legacy system cannot run in your environment, the best possible verdict is PARTLY PROVEN. The video does not mention that ceiling. The canary shows that the test suite can fail. It does not show that the suite covers the legacy behavior, and the invented-input comparison is the control aimed at that second question.

## What the video does not show

The speaker only reads the README. At the sampled frame, and in the transcript, there is no run of the plugin on a real legacy codebase, no PROVEN or NOT PROVEN verdict and no measured failure rate. The closing line "seems to be very well thought out from what I'm reading here" (67.70 s) is an explicit reading, not a test. The README has a results table (the commit title mentions it), but the video does not cite it. The description's claim that this is how "most legacy rewrites go wrong without anyone noticing" is the speaker's opinion. I did not find a source for the "most" in the README's wording, so I treat it as unverified.

## Key Takeaways

- **Verified:** The plugin exists at plugins/code-modernization in anthropics/claude-plugins-official, and its README lists the four deliverables, the 10 invented inputs, the canary and the one-byte comparison check as the video says.
- **Partial correction:** The video says at 4.50 s the plugin "just dropped". Frame 1 shows the last commit as "3 weeks ago", and the commit history lists 35 commits dated Sep 25, 2026 and a 1.0.0 version, all before the 14 November post.
- **Clarification:** "Compares every byte" applies after masking fields that must vary, which the README lists, and floating-point output is compared within a declared tolerance.
- **Context:** The best verdict available when the legacy system cannot run is PARTLY PROVEN, which the video leaves out.
- **Unverified:** The description's statement that most legacy rewrites fail unnoticed has no source in the video or the README. The video shows no live run of the plugin.

## Resources

- [code-modernization README, anthropics/claude-plugins-official, GitHub](https://github.com/anthropics/claude-plugins-official/blob/main/plugins/code-modernization/README.md) establishes the four deliverables, the verification controls, the canary, and the PROVEN, PARTLY PROVEN and NOT PROVEN verdicts I cite (fetched through a summarizer, so my wording is paraphrase).
- [Commit history for plugins/code-modernization, GitHub](https://github.com/anthropics/claude-plugins-official/commits/main/plugins/code-modernization) establishes the 35 commit titles, the Sep 25, 2026 date header, the author morganl-ant and the 1.0.0 version I cite (also via summarizer, no per-commit timestamps).
- [Plugins directory, anthropics/claude-plugins-official, GitHub](https://github.com/anthropics/claude-plugins-official/tree/main/plugins) establishes that a code-modernization folder exists in the repo.
- The linked YouTube full video (https://youtu.be/2MuKFVGpJ-k) was not fetched, and the staged links folder was not available, so I cite nothing from it.
