Claude Code Builds Cross-Platform Task Manager From 107-Page Spec

A Task Manager reboot that started in a hospital waiting room

Dave W. Plummer wrote the original 80KB Windows Task Manager, which first shipped publicly in Windows NT 4.0 in 1996. He has now released TMOG (Task Manager OG), a new system monitor that runs natively on Windows 11, macOS and Linux.1 He built it largely with Anthropic's Claude Code. He did not give it a one-line prompt. He gave it a 107-page specification.12

The circumstances explain the timing. Plummer did much of the work in a hospital while his son had an emergency appendectomy. He had a rough working version by the time his son was first discharged.13 When his son was readmitted, Plummer got another long stretch of time, and he spent it on features.4

The rollout came in stages. A macOS beta came out on July 31, and Windows and Linux builds followed on August 23.79 A third update and the first paid Pro releases arrived on September 7.9

The spec is the story

The project began as an argument over whether someone could vibe code Microsoft Word today. Plummer doubted it and picked a smaller target he knew well.23 He put the main lesson plainly: you can tell Claude to "write a task manager," but the result will be "hot garbage."2 He compared his 107-page document to something written by a caffeinated 1990s product manager. That is an ironic choice for an engineer known for squeezing the original tool down to about 80KB.2

The coverage broadly agrees on what happened next. Plummer gave Claude Code the spec, and on a MacBook Air M5 he had first builds within about four to four and a half hours. He says they worked the first time, and refinement followed.146 Some later write-ups describe the spec in detail: data structures, metric formulas and UI states, all defined before any code was generated.10

The reports also agree on how to read this, and that reading is the right one. TMOG does not show that AI can replace a systems engineer. It shows that an expert's knowledge, written down carefully enough, can be turned into code very quickly. One analysis said Claude Code supplied speed, while Plummer supplied the architecture, the requirements and the ability to spot when the generated code was wrong.7 Another noted that the more interesting lesson is how much still had to be spelled out. Three operating systems handle processes, services, hardware counters, permissions and interface conventions in very different ways.5

The "four hours" headline also needs context. Plummer says the project took under six weeks, with two collaborators and Claude Code doing the heavy lifting.29 The fast first build was the start of the work, not the end.

Native, not Electron

The architecture is probably the most ambitious part. Many cross-platform apps put one web app inside Electron and ship it everywhere.2 TMOG instead has a separate native interface for each platform: Swift and AppKit on macOS, Qt 6 on Linux, and native Windows code. All three sit on a shared C++ core that keeps metric definitions, actions, models and serialization consistent.79

That choice means much more platform-specific code. This is exactly the kind of tedious but well-defined work coding agents handle well when the requirements are clear. For a small team, the approach makes three native front ends realistic in a way they might not otherwise be.

Features include meters animated at 60Hz instead of once per second, separate colours for efficiency cores, and retro phosphor themes in green, amber and blue.2 Twelve views cover processes, startup apps, services, power and frequency, network connections, installed applications, disk space and benchmarks.9 The free version covers live monitoring on all three platforms. A $39.95 Pro tier adds Flight Recorder, which records system activity so users can replay a slowdown after it happens.210

Where the coverage diverges

The reports differ on several technical details, which is unsurprising for a beta that changes quickly. One outlet says the Windows build uses Win32 controls.7 Others say it renders with Direct2D.910 Several reports list macOS 13 as the minimum.57 A later guide says macOS 14 or newer.10 Most coverage in late August described version 0.1.1 beta.47 The later guide refers to a 1.0.0 free tier.10 Code signing also differs. Earlier reports said the downloads were not digitally signed, so users might have to approve them manually.57 The September guide tells users to check the signature.10

The best explanation is that TMOG has changed between articles, not that any outlet got it wrong. Readers should treat these numbers as snapshots and check the official site before installing. The signing question matters most. A utility that needs deep access to processes and hardware should be signed, and early users had to deal with operating-system warnings.7

Apple's rules have also shaped how TMOG is distributed. Plummer reportedly planned a Mac App Store beta. But the sandbox blocked the process access a task manager needs, and a sandboxed build reportedly showed zero processes.5 Direct download was the only practical option.

The trust question vibe coding can't dodge

The launch prompted the usual argument about AI-generated code. Commenters said running a vibe-coded app is reckless because nobody knows what vulnerabilities it contains. Others replied that the author is the person who wrote the original tool.1 Plummer reportedly pointed to his long experience and said he also used AI to look for security flaws.6 Early users are still finding bugs, and Plummer has been answering individual comments himself.12

This is where AI code review becomes relevant. Writing code with AI is no longer the hard part. Checking it is. Anthropic seems to agree, judging by recent Claude Code releases. The October changelogs include several fixes to the Code Review feature. Blocking comments no longer appear with a contradictory "nit" label. Finding comments no longer cut off mid-sentence.11 Version 2.1.287 added Claude Mods, a way for plugins to change deeper behaviour. It also added an optional built-in mod called "You should know," where a second agent watches the session and flags things the user or Claude might miss.1215

The same few weeks were dominated by security fixes. Releases from 2.1.289 to 2.1.292 closed gaps where permission rules were being skipped. One fix, for file reads through UNC paths, was labelled a security issue, and developers were told to update to 2.1.296 or later.14 Version 2.1.295 made hooks fail closed: if a hook breaks, the action is blocked instead of allowed.17 Version 2.1.296 lets teams pin every subagent in a workflow to one chosen model. It also fixed a bug where a hook that denied a tool call did not actually end the agent's turn.18

The pace is notable. Trackers count more than 400 Claude Code releases.1319 Recent model changes made Opus 5.5 the default Opus model on September 22 and Sonnet 5.5 the default Sonnet on September 28.16 Claude Haiku 5.5 then became the default Haiku on the API, with a 1M-token context window.14 Competition is pushing the speed too: OpenAI's Codex CLI shipped an update within 48 hours of Claude Code 2.1.296.18

What TMOG actually proves

Across the coverage, TMOG reads best as a well-run engineering project in which the coding agent served as a fast and capable implementer. One report called it a credible test of agentic coding directed by an expert.7 That is a fair description. Plummer still had to decide what a task manager should measure, how each operating system exposes that data, and what 60Hz animation means for an app that is measuring the load it adds itself. Claude Code turned those decisions into three native programs faster than a small team could have done alone.

The lesson for developers is not that anyone can now build anything. Detailed specifications are becoming the main input to software, and review tools, permission controls and fail-closed hooks are becoming the necessary safeguards. Plummer's hospital-room experiment worked because he already knew the answers before he asked. Whether less experienced developers can match that depends more on the review tools Anthropic and its rivals are now racing to improve than on generation speed.