March 20, 2026super-user

Super User Daily: Mar 21, 2026

Claude Code Channels launched — control your coding agent from Telegram and Discord. Google shipped DESIGN.md connecting Stitch to Claude Code and Cursor. And someone ran a full Pentagon budget audit with Claude Code, flagging $4.2B in potential savings.
@trq212 [Claude Code]
Claude Code#1
https://x.com/trq212/status/2034642575819170174
Anthropic launched Claude Code Channels, allowing users to control their Claude Code sessions through Telegram and Discord MCPs. This means you can message your coding agent directly from your phone through your existing messaging apps, removing the need to be at your terminal. A major step toward always-on agent accessibility.
@PawelHuryn [Claude Code]
Claude Code#2
https://x.com/PawelHuryn/status/2034551610551610000
Google shipped DESIGN.md — a portable, agent-readable design system file. Stitch now has an MCP server connecting directly to Claude Code, Cursor, and Gemini CLI. The pipeline works today: a PM describes the business objective, Stitch generates the UI, the coding agent reads DESIGN.md and builds against it. PRD to design to code used to be three teams and three handoffs — now it's one loop with one context file.
@milesdeutscher [Claude Code]
Claude Code#3
https://x.com/milesdeutscher/status/2034555767555599000
Someone ran a full audit on the Pentagon's budget using Claude Code. The results: 340 flagged items, $4.2 billion in potential undercuts, and overspending by 10x on certain line items. When institutions refuse to audit themselves, AI agents become the auditors. A powerful non-coding application demonstrating Claude Code's analytical capabilities on massive datasets.
@qkl2058 [Claude Code]
Claude Code#4
https://x.com/qkl2058/status/2034531153633000000
A Chinese MIT student built a physical trading box — book-sized, four buttons (buy, sell, confirm, cancel), one microphone. He speaks commands like "scan Premier League odds," Claude Code analyzes the data, and the device executes. Starting from his $2,400/month dorm, the system has generated $4.5 million in 8 months across 4,548 predictions on Polymarket. His robotics professor gave him a B+ with the note "interesting idea, limited practical value."
@NainsiDwiv50980 [Claude Code]
Claude Code#5
https://x.com/NainsiDwiv50980/status/2034529850000000000
A developer ran a ~400 billion parameter AI model (Qwen3.5 397B) on a 48GB MacBook — no cloud, no data center. They fed Claude Code with Karpathy's autoresearch repo and Apple's "LLM in a Flash" paper, then Claude Code figured out the implementation. Result: ~1 token/sec, ~21GB RAM, rest streamed from SSD. We're entering a world where frontier-scale models run on consumer hardware.
@Emarky [Claude Code]
Claude Code#6
https://x.com/Emarky/status/2034519055000000000
A creator had an entire video edited by Claude Code — just gave it the files, and Claude handled the cuts, transitions, and assembly. No video editing software, no manual timeline work. This extends Claude Code's reach beyond coding into professional content production workflows.
@Maxwell_SCU [Claude Code]
Claude Code#7
https://x.com/Maxwell_SCU/status/2034636130734000000
Instead of using traditional torrent clients, a user had Claude Code write a custom magnet link downloader optimized for their specific network conditions. Claude Code analyzed network configuration and parameters, then auto-tuned download settings. The result was dramatically faster download speeds than off-the-shelf tools — a practical example of AI-generated software tailored to individual infrastructure.
@garrytan [Claude Code]
Claude Code#8
https://x.com/garrytan/status/2034532101626000000
Y Combinator president Garry Tan added a /codex skill to Claude Code that performs plan and code review. The skill has found significantly more bugs and helped produce more elegant architectural designs. He's also launching native Codex support so his GStack tool can run on Codex — demonstrating how power users combine multiple AI coding tools for complementary strengths.
@AlexFinn [OpenClaw]
OpenClaw#9
https://x.com/AlexFinn/status/2034113806266310000
A single prompt sped up an OpenClaw instance by 95%. The issue: every cron job's output was being loaded into context, accumulating months of data. The fix was simply checking and cleaning ~/.openclaw/agents/main/sessions/ and sessions.json. A critical maintenance tip for anyone running OpenClaw long-term.
@jordymaui [OpenClaw]
OpenClaw#10
https://x.com/jordymaui/status/2034549593108598000
A user's OpenClaw agent has been generating real revenue — selling to other agents autonomously, with zero human intervention. The owner reports doing nothing while the agent conducts agent-to-agent commerce. An early example of autonomous AI agent economies actually producing income.
@jshchnz [Claude Code]
Claude Code#11
https://x.com/jshchnz/status/2034368892144366000
Survey from SF's Agents Anonymous meetup: 90% use Claude Code, 60% Codex, 30% Cursor. 80% have prompted a coding agent from mobile. 50% have not handwritten a single line of code this year. 99% believe they're more productive with agentic coding agents. The developer workflow has fundamentally shifted.
@wesbos [Claude Code]
#12
https://x.com/wesbos/status/2034411125297248000
Wes Bos ran an unscientific comparison: build a Twitter clone with the same stack (Better Auth, Vite, SQLite, Drizzle, TypeScript, React). Composer: 5 min, $6.04, 1,250 LOC. Opus: 19 min, $10.43, 1,000 LOC. GPT: 22 min, $14.15, 2,000 LOC. Composer was fastest and cheapest, but Opus used caching more aggressively, potentially saving on repeated tasks.
@sogitani_baigie [Claude Code]
Claude Code#13
https://x.com/sogitani_baigie/status/2034561100269000000
A non-engineer built a complete talent management system using Claude Code, documenting the entire process in a tutorial. Zero lines of code written manually. The tutorial avoids technical jargon entirely, proving that business users can build production internal tools with Claude Code alone.
🗣 User Voice
User Voice:

1. Rate limits and account bans remain the top frustration. Users report Claude Code sessions dying mid-work, forcing workarounds like routing through AWS Kiro or using proxy tools. (@plantegg, @sersweai)

2. Battery drain is severe — Claude Code on MacBook Pro M4 drains 100% to 0% in 3 hours even when idle. Users are moving to VPS-based workflows to cope. (@levelsio)

3. Demand for mobile-native Claude Code control is intense. Channels (Telegram/Discord) partially addresses this, but users want direct iOS/Android apps. (@jamischarles, @trq212)

4. Context window management is a recurring pain point. Long-running sessions accumulate stale context that degrades output quality. Users want better automatic context pruning. (@AlexFinn)

5. OpenClaw's install complexity contradicts its "no-code" promise. Non-technical users find setup difficult despite marketing that suggests otherwise. (@romi_hoshino)
📡 Eco Products Radar
Eco Products Radar:
Claude Code (257+ mentions) — Anthropic's terminal AI coding agent
OpenClaw (267+ mentions) — Open-source autonomous AI agent framework
Cursor (30+ mentions) — AI-native code editor
Codex (25+ mentions) — OpenAI's coding agent
Stitch (15+ mentions) — Google's AI-native design tool with DESIGN.md
Gemini (12+ mentions) — Google's AI model family
Telegram (8+ mentions) — Messaging platform, now a Claude Code channel
Obsidian (6+ mentions) — Local markdown knowledge base
Vercel (5+ mentions) — Frontend deployment with agent skills
n8n (4+ mentions) — Open-source workflow automation
Kiro (3+ mentions) — AWS coding IDE with Claude models
← Previous
GitHub Stars Daily Spotlight — March 21, 2026
Next →
Ideas Radar: Mar 21, 2026
← Back to all articles
>_