<- all tokdocs
2026.09.13@codenameposhan

The Code Modernization Plugin in This TikTok Is Real and Five Months Old, and the Tour Skips Its Human Approval Gate

Coding
Watch on TikTok

View on TikTok ->

Every command name, install string, and workflow claim in this video checks out against Anthropic's own repo, but the plugin landed there on April 24, 2026, and the four-command tour leaves out the step that actually makes the workflow controlled. The clip runs 1 minute 57 seconds at 1080x1920. It opens on a screen recording of the plugin README at github.com/anthropics/claude-plugins-official, cuts to a dark motion-graphics sequence with an orange starburst and monospaced labels like STRUCTURED WORKFLOW and NOTHING WAS READ, then spends its middle on a real terminal session running inside DevSwarm, and closes with about 20 seconds of DevSwarm product footage. Burned-in captions run the whole length. The music is credited to "Epic News" by Audiodeity. At the time I pulled the metadata the video had 86,300 views, 3,430 likes, 3,877 saves, 740 reposts, and 61 comments.

What the run on screen actually did

The terminal footage is not a mock-up. The header reads Claude Code v2.1.266, Opus 5 (1M context), Claude Max, working directory ~/Documents/R&D_Projs/calculator. After /plugin install code-modernization@claude-plugins-official, a /reload-plugins line reports "Reloaded: 3 plugins · 178 skills · 15 agents · 2 hooks · 1 plugin MCP server · 0 plugin LSP servers." The install confirmation panel in the Discover tab lists the plugin as "By: Anthropic," "Last updated: Jul 8, 2026," and gives a context cost of roughly 1.2k tokens every turn and roughly 40.6k tokens when invoked. Those two token figures come from the on-screen panel and I could not verify them independently, since the number depends on the installed build.

/modernize-preflight opens with what the plugin calls Check 0, five questions it asks the operator because the source cannot answer them: scope, whether you can build and test locally, bespoke build infrastructure, prior modernization attempts, and what is off limits. It then runs eleven shell commands and prints a per-check verdict table. Tests come back red with "No tests, no framework, no CI." Telemetry comes back red. VCS history comes back amber with "Exactly one commit, no remote," which the report notes disables change-frequency risk ranking. Path convention comes back amber because there is no legacy/ directory. The smoke tests pass: cobc -fsyntax-only rc 0; javac rc 0; --release 8 rc 0. Preflight finished in 2 minutes 45 seconds at 11:18 AM.

/modernize-assess launches three background agents by name, legacy-analyst twice for a structural domain map and a technical debt scan, plus security-auditor for a security scan. It finished at 11:26 AM after 7 minutes. /modernize-map produces TOPOLOGY.html, an interactive circle-pack the video scrolls through. Its header reads "calculator · 64 modules · 3 data stores · 121 edges · 5 entry points," with edge-type toggles for calls, dynamic dispatch, reads, and writes, and a "Business flow walkthrough" dropdown. The artifact table printed afterward describes topology.json as "67 nodes, 121 edges, 4 persona flows." Those two figures agree: 64 modules plus 3 data stores is 67 nodes.

The README is quoted verbatim, and the commit history dates the plugin

I fetched the raw README. The video's opening shot shows it unedited, including the pipeline line preflight → assess → map → extract-rules → brief → (reimagine | transform | uplift) and the three method bullets for transform, reimagine, and uplift. The creator's paraphrase in the voiceover tracks the source closely. The install string /plugin install code-modernization@claude-plugins-official is exactly what the README publishes.

The repo is real and it is Anthropic's. anthropics/claude-plugins-official was created on November 20, 2025, is Apache 2.0 licensed, and describes itself as the "Official, Anthropic-managed directory of high quality Claude Code Plugins." When I queried the GitHub API on September 15, 2026 it showed 36,309 stars, 4,077 forks, and 1,038 open issues, with 39 plugins in the plugins/ directory. The video's browser chrome shows 970 issues and 69 pull requests, consistent with a snapshot taken a few days earlier. The plugin manifest lists the author as Anthropic with the address support@anthropic.com, and the listing at claude.com/plugins/code-modernization carries an "Anthropic Verified" badge and showed 5,860 installs when I checked.

The timing is where the video overstates. The path plugins/code-modernization has 20 commits, all by the same GitHub account, morganl-ant. The first is bdca23e, "Add code-modernization plugin," dated April 24, 2026. The most recent is 96e3e8c, "code-modernization: pilot-first uplift migration, brief-driven execution, deeper preflight," dated July 9, 2026. That second commit is the one visible in the blame bar in the video's own screen recording, labeled "2 months ago." So the plugin was roughly five months old when the video described it as something Anthropic released, and it had not been touched in two months. Nothing in the video is false about authorship. The framing implies a launch that did not happen in September.

The history also shows the plugin was tested against something substantial. A May 11, 2026 commit reads "Harden code-modernization plugin from a real CardDemo dry run," and the topology screenshot in the README is captioned as AWS CardDemo. That is aws-samples/aws-mainframe-modernization-carddemo, a public Apache 2.0 COBOL credit card application with 275 stars, built by AWS to demo mainframe migration patterns.

The three commands the tour skips

The video walks through preflight, assess, and map, then jumps straight to the three build methods. The published pipeline has three more steps in between and after, and one of them is the thing the video's own thesis depends on.

/modernize-extract-rules mines calculations, validations, eligibility logic, and state transitions into Given/When/Then "Rule Cards" with file:line citations and confidence ratings, producing BUSINESS_RULES.md and DATA_OBJECTS.md. /modernize-brief is the one that matters most. It synthesizes the discovery artifacts into a phased modernization brief, refuses to run if any discovery artifact is missing, and enters plan mode as a human-in-the-loop approval gate. The README states that the execution commands read the brief and treat each phase's entry criteria as gates, so editing the brief steers execution. The voiceover says Claude "first tries to understand the legacy system, document what it found, and then modernize it in a controlled workflow." The brief is the control. Leaving it out of a tour built around that claim is the video's largest gap.

/modernize-harden runs a security pass on the still-running legacy system and writes SECURITY_FINDINGS.md plus a security_remediation.patch that you apply yourself. The README says it never edits legacy/. The repo confirms the shape of all of this: 10 command files, 8 agent files, and 6 JavaScript workflow scripts. The names in the install panel on screen match the repo file names one for one, all 10 commands and all 8 agents.

Two more things the video never mentions. The README recommends a .claude/settings.json that denies Edit(legacy/**) and Write(legacy/**) outright, which is the only real enforcement of the "legacy stays untouched" promise the animated overlay puts on screen. And the README carries a safety section stating that analyzed code is untrusted input, since a hostile codebase can plant comments like "ignore previous instructions" to steer what lands in BUSINESS_RULES.md. Discovered credentials are masked as AKIA**** and quarantined in a gitignored SECRETS.local.md. The README also advises anyone who ran an early version on a real system to check whether analysis/ artifacts were committed and rotate anything exposed. That is a live warning aimed at the exact audience the video is pitching.

The calculator is three languages, and assess found a real defect

"This legacy calculator app" undersells what is on screen. The assess output prints a file inventory: index.html in HTML 4.01 plus ES5 at 531 total lines, Calculator.java in what it labels Java 1.4 idiom at 442 lines, calc.cbl in COBOL-85 at 423 lines, a GNU Makefile at 50 lines, and a 62-line README. Three separate implementations of a calculator, in three languages, sharing one on-disk file.

The assessment's headline finding is worth more than the plugin tour around it: "the three implementations are not the same calculator." COBOL computes one A op B per invocation. Java and the browser both keep an accumulator and support chained operations, so 1 + 2 + 3 means different things depending on which binary you run. Number formatting diverged too. The report shows COBOL writing PIC S9(11)V9(4) fixed point, rendering 12.0000 + 5.0000 = 17.0000, while Java uses DecimalFormat("0.####") and renders 9 + 9 = 18. Both formats are already sitting in the live CALCHIST.DAT history file with nothing marking which engine wrote which record. The report's conclusion is blunt: any plan that assumes "extract the rules once, they're the same in all three" is wrong before it starts.

It also found a genuine bug. COBOL writes overflowed results to the permanent history as 0.0000 with no marker, cited to calc.cbl:208-238, and the operand parse feeding it has no overflow trap, so the audit trail can contain confident-looking garbage. Security came back clean, with five low-severity shared-file findings covering symlink following, no locking, and default permissions, none exploitable in the documented single-user case.

The video's uplift framing does not match what the demo did. The assessment's recommendation on screen is "Rebuild → /modernize-reimagine, at 1.4 KSLOC with both legacy runtimes building and running here, so true dual-run equivalence testing is available." The operator then typed modernize the app now in plain English rather than any slash command, answered three questions (TypeScript monorepo, chained arithmetic, migrate both legacy history formats), and the session built a zero-dependency TypeScript monorepo using Node 24 native type stripping and node:test. Test counts climb on screen from 13/13 to 50 to 67/67. The last verification pass reports chained arithmetic working, a divide-by-zero refusal shown in red, and history correctly not recording the failed calculation. One background command, "Start web server," failed with exit code 144 along the way. The run timer at the end reads 14 minutes 39 seconds and 64.8k tokens.

The COCOMO figure and the .NET 8 example

The assess output prints COCOMO-II index: KSLOC 1.446 → 2.94 × 1.446^1.10 = 4.41. The constant and exponent are correct. COCOMO II's post-architecture effort equation is PM = A × Size^E, with A = 2.94 in the COCOMO II.2000 calibration and a nominal E of 1.0997, which rounds to the 1.10 on screen. The arithmetic is right as well: 1.446^1.10 is about 1.5005, times 2.94 gives 4.41.

The size input is the part I would question. Adding the four source files' total line counts gives 531 + 442 + 423 + 50 = 1,446, which is exactly the 1.446 KSLOC used. That is physical lines including blanks and comments. COCOMO II defines size in logical source statements per the SEI definition checklist, which excludes blank lines and comment lines. Using the code column instead gives 1,187 lines, and 2.94 × 1.187^1.10 comes to 3.55 rather than 4.41, a difference of about 24 percent. The plugin's own README defuses most of this, since it states that the COCOMO figure is used only as a relative complexity index for ranking and sequencing systems and never as a timeline or cost, and a June 9, 2026 commit is titled "COCOMO is a complexity index, never a modernization timeline." A consistently inflated index still ranks systems in the same order. It is worth knowing the number is not a standard COCOMO size.

The .NET example is the one piece of advice in the video that has an expiry date. The voiceover says that moving from an older .NET version to .NET 8 calls for the uplift workflow, which mirrors the README's own example of .NET Framework 4.8 to .NET 8. Microsoft's support policy puts .NET 8 in maintenance, with end of support on November 10, 2026, less than two months after this video posted. .NET 10 shipped on November 11, 2025 as the current LTS release and is supported through November 14, 2028. Anyone starting a .NET Framework uplift now should be targeting .NET 10. The README example is stale, and the video repeated it without checking.

The DevSwarm slot at the end

The last 20 seconds are a product pitch, and the specifics hold up. DevSwarm is a real product from 21st Idea, Inc., with a public repo at devswarm-ai/devswarm that had 132 stars when I checked. The claim that it runs multiple projects with multiple branches in parallel, each branch in its own isolated environment, matches the product's own description of git worktrees paired with an assistant terminal and an editor. "Try it free" is accurate: there is a free ad-supported tier, with Pro at $8 per month and Team at $18 per month on the pricing page.

One number I could check directly against the footage. The DevSwarm pricing page advertises 19 coding agent options. Frame 57 shows the agent panel with "Open installed agents (8)" listing Claude Code, OpenAI Codex, Amp, Goose, Cursor CLI, Aider, Open Code, and Droid, and "Not installed (11)" listing Gemini, Copilot, Qwen Code, Auggie, Mistral Vibe, Amazon Q, Atlassian Rovo, Crush, Cline, Plandex, and Qodo. That is 19, matching the marketing claim exactly.

The final frame is an accidentally honest one for a promo. The DevSwarm workspace shows a Codex pane returning Warning: Model metadata for gpt-5.4 not found followed by a 400 error reading The 'gpt-5.4' model is not supported when using Codex with a ChatGPT account, sitting directly under the "try it free at devswarm.ai" caption.

Key Takeaways

  • The plugin is real, Anthropic-authored, and Apache 2.0. I fetched the raw README and it matches the video's screen recording word for word, including the install string /plugin install code-modernization@claude-plugins-official.
  • The "released" framing is off by five months. The first commit adding plugins/code-modernization is dated April 24, 2026, and the most recent of its 20 commits is July 9, 2026. The video posted September 13, 2026.
  • The tour covers preflight, assess, and map, then skips extract-rules, brief, and harden. /modernize-brief is the human-in-the-loop approval gate that refuses to run without the discovery artifacts, and the execution commands read it as their phase gates. It is the mechanism behind the video's own "controlled workflow" claim.
  • The command and agent lists shown in the install panel match the repo exactly: 10 command files and 8 agent files, plus 6 workflow scripts the video never mentions.
  • The COCOMO II constants on screen are correct. A = 2.94 and nominal E = 1.0997 are the published COCOMO II.2000 values, and 2.94 × 1.446^1.10 = 4.41 is accurate arithmetic. The 1.446 KSLOC input is total physical lines including blanks and comments, which is not COCOMO II's logical SLOC definition. Code-only lines give 1.187 KSLOC and an index of 3.55.
  • The .NET 8 uplift example is dated. .NET 8 reaches end of support on November 10, 2026. .NET 10 has been the current LTS since November 11, 2025 and is supported through November 14, 2028.
  • The demo never ran transform, reimagine, or uplift as a command. The assessment recommended /modernize-reimagine, and the operator then typed "modernize the app now" in plain English.
  • The most useful thing in the whole clip is an assess finding the voiceover never mentions: the three calculator implementations disagree on chained arithmetic and on number formatting, and both formats are already mixed together in one live history file.
  • DevSwarm's claims check out. The free tier exists, Pro is $8 per month, and the 19 supported agents advertised on the pricing page match the 8 installed plus 11 uninstalled entries visible in the video's own agent panel.
  • I could not independently verify the plugin's on-screen context cost of roughly 1.2k tokens per turn and 40.6k tokens when invoked, since that figure comes from the installed build rather than the repo.

Resources

Published September 13, 2026. Writeup generated from a favorited TikTok.