App builders report AI-assisted shipping friction in CSS-to-Tailwind refactors

• A CSS-to-Tailwind migration still got stuck in agent babysitting, with thekitze's refactor note saying he had to resume it five times.
• Local dev split into camps: thekitze said he has not used localhost all year in his localhost post, while Arch users were still relying on ports after OpenAI's Linux desktop preview, according to LLMJunky's Arch reply.
• Verification became the shared pain: bcherny's code-review post said LLM bugs have shifted toward system design, UI usability, and missing broader context.
• Creative wins still looked real, with MengTo's three.js tower site coming in under 1 to 2 MB and levelsio's finance dashboard built from exported bank CSVs.
• Agent UX is still leaking complexity, from hidden model and reasoning details in LLMJunky's subagent rant to ChatGPT's messy Chat, Work, and Codex split in Peter Yang's onboarding complaint.

OpenAI's Linux preview post lists Ubuntu 24.04 and 26.04, Debian 13, Fedora 43 and 44, x64 and ARM64, and .deb and .rpm packages. Anthropic's Claude Code v2.1.228 changelog includes fixes for dead redraws, Git Bash discovery on Windows, cross-session messaging inboxes, and Remote Control /resume leakage. A same-week worktrees guide framed isolated folders as a way to stop parallel Claude Code sessions from reading each other's half-edited files.

The failed refactor was mundane: move a codebase from .css to Tailwind. thekitze said the agent needed five resumes to finish it.

On Reddit, a founding designer's complaint described a similar fatigue pattern: no visible GUI-style error state, ignored instructions, wasted tokens, and confident outputs that still need policing.

bentossell described the less angry version of the same work: in his agent-learning post, he compared following agents to reading a book full of unfamiliar words, then stopping mid-chat for questions until the concepts click.

AIandDesign said an older app build with Sonnet 3.5, Sonnet 3.7, and Augment Code took many months, with his reply pointing to the resulting iOS game.

Agents are now good enough to expose the backlog. thekitze asked a clanker to list every app he had built across all his machines and came away with “so much cooking so lil shipping.”

levelsio's Claude Code dashboard started with exported CSVs from personal, business, and brokerage accounts. The useful part was the queryable dataset after the cleanup, not the chart.

His findings included:

• forgotten subscriptions running for two years
• 30% withholding tax on US stock dividends
• roughly $400K spent on domains
• about $24K per month in AI GPU costs for Photo AI
• business costs of about $5K per month without AI, rising to about $28K per month with AI

karenxcheng's Grok Bot setup used the bot for a package tracker, a client-footage handoff to a video editor, and a vacation-mode email router. In her UI note, she said the interface felt closer to iMessage than an AI coding tool.

thekitze said he had not used localhost:port all year, then separately posted “you will regret remote coding.” That pair captures the weird middle of 2026 app-building: local servers feel old to some builders, but remote machines still produce dread.

ericzakariasson's remote-machine reply put one reason plainly: he never liked the developer experience of worktrees and had already moved to a remote machine or cloud agent.

Speed remained a drag. Everlier's agent-speed complaint said agents are still too slow.

OpenAI's Linux preview brought ChatGPT, Work, and Codex into one desktop app on supported systems, according to the OpenAI community announcement. The official list skipped Arch.

LLMJunky's Arch follow-up called the release a step in the right direction while saying Arch users still needed a port. In another reply, he said the port could probably switch to the new app, though he doubted it would change much.

The harder Linux limit was computer use. LLMJunky's Wayland warning said Linux computer use on Wayland is worse than people expect because Wayland is too restrictive for the required infrastructure.

thekitze said he never trusts agent web browsing and still gets better results by opening three tabs himself. The suspected causes in his post included captchas or blocked access, but the complaint was about the outcome: worse search than manual browsing.

Opacity showed up inside agent UIs too. LLMJunky's file-editing reply said Claude shows an “incredibly undetailed overview” of the files it is editing, while his subagent screenshot asked to see each subagent's model and reasoning level.

Peter Yang's ChatGPT desktop complaint came from onboarding his parents: Chat, Work, and Codex were separated, and web, desktop, and mobile behaved inconsistently. In the same trust bucket, levelsio's Claude refusal post said Claude had become preachy and that he would jump to Grok if it became good enough for coding.

bcherny said coding-agent bugs have changed: fewer off-by-ones, more system design, UI usability, and missing-context failures. His suggested pattern was adversarial code review, including Claude's built-in /code-review modes.

LinusEkenstam's Sonar survey post put numbers on the trust gap: 42% of committed code was AI-generated, 96% of developers did not fully trust it, and 48% always checked it before committing.

bcherny's low-effort reply said /code-review low can cost under one cent. His forked-agent reply said /code-review had used a forked agent since version 2.1.218, then a later bug reply said a specific issue had been found and a fix was incoming.

nicklaunches put the job change in one reply: reviewing, not writing. In another reply, he said blind accepting eventually bites.

Worktrees solved one class of agent collision and created their own cleanup tax. tequilafunks asked whether creating a worktree for every task triggers antivirus on Windows.

bcherny's stale-worktree reply asked whether Claude Code should build in a cleanup loop for stale worktrees. The Rift repo linked from tequilafunks framed copy-on-write workspaces as a faster, space-efficient alternative to git worktrees, with Linux btrfs snapshots, macOS APFS clonefile support, and no Windows workspace creation yet.

The most concrete creator wins came from workflows that gave the agent a structured surface to modify.

MengTo's interactive three.js site used 3D instead of huge static images and stayed under 1 to 2 MB. The site could generate themed towers, weather, and lighting variations after the foundation was in place.

The build stack was explicit:

• Claude Code for the harness
• Higgsfield for textures, sound effects, and music
• GPT image 2 for texture images applied to the 3D objects, according to MengTo's texture reply
• one HTML file, according to MengTo's file reply

DannyLimanseta's Grok Bot run said the bot generated, cleaned, cropped, and integrated 74 game art assets in about two hours. The same report also covered playtesting, itch.io build uploads, Figma MCP wireframes from a PRD, subscription audits, and newsletter unsubscribes.

Adham Dannaway's Bolt template demo showed a budget-tracker template customized by prompt, and Bolt's template marketplace framed templates as free starts for portfolios, landing pages, internal tools, MVPs, dashboards, and ecommerce. jakubkrehel's /interface-review skill added a review pass across UI, typography, layout, color, writing, and accessibility, while Creative Dash's Forge canvas demo rebuilt a node-grid canvas with performance gains.