Super User Daily: April 09, 2026
April 7 was dominated by two forces pulling in opposite directions. On one side, builders are wiring Claude Code into every corner of their lives, from recruiting pipelines to overnight ad generation to reverse-engineering network infrastructure via SSH. On the other, the community is loudly demanding transparency on token consumption and expressing frustration with perceived quality regressions. The non-coding use cases are where things get genuinely surprising.
@om_patel5 [Claude Code]
https://x.com/om_patel5/status/2041332038276428177
Someone connected Claude Voice Mode to Claude Code using Apple Reminders as a bridge. You talk into your AirPods while walking, the voice mode drops the prompt into a Reminders list, and Claude Code checks every 60 seconds to pick up the task. The code writes itself on his computer at home while he is outside brainstorming. He is not a developer and vibe coded the entire bridge system himself. This is hands-free coding with zero screen time.
@PawelHuryn [Claude Code]
https://x.com/PawelHuryn/status/2041595776074236328
Built a local dashboard that reads the JSONL files Claude Code already writes to ~/.claude/projects/. Every session, every turn, every token is logged there but Anthropic gives you nothing except a progress bar saying "63% used." His numbers over 30 days: 440 sessions, 18,000 turns, $1,588 in API-equivalent costs. The dashboard scans those files, builds a SQLite database, and serves charts on localhost:8080. Zero dependencies, Python standard library only. Open source, MIT license.
@mikefutia [Claude Code]
https://x.com/mikefutia/status/2041567383056519304
Built a Claude Code skill that ships 50 static ad concepts to his desktop every morning. You feed it your customer reviews, winning ads, and top comments. It studies what hooks and pain points are converting, pulls from 15 proven direct response templates, writes 50 new concepts in your brand voice, fires prompts to Nano Banana 2 for finished images, and drops everything into a dated folder ready to upload to your CBO. No briefing designers, no 3-day turnarounds.
@kenn [Claude Code]
https://x.com/kenn/status/2041438528090058950
Shared a 25-step multi-AI collaboration workflow for implementing complex features. Start in Codex Plan mode on High, let it ask 5-15 clarifying questions, then export the plan to a docs folder. Open Claude Code to review it since Claude catches over-engineering that Codex tends toward. Copy Claude feedback back to Codex, iterate until convergence. Then hand off to Cursor Composer for fast implementation. Use staged vs unstaged git diffs to control what each AI sees. Codex is the strategist, Cursor the implementer, Claude the cleanup crew.
@m0370 [Claude Code]
https://x.com/m0370/status/2041467668625301738
Japanese researcher who has Claude connected to Gmail and Google Calendar via MCP, automatically reading email every morning at 5 AM and organizing reminders. Fed his entire Obsidian vault and years of PowerPoint presentations to Claude Code to draft next lecture slides. Also has it reading a Japanese hotel booking site to plan conference trip schedules. Says the point of no return has been crossed for personal data and AI.
@kazuuuya_tk [Claude Code]
https://x.com/kazuuuya_tk/status/2041423341643612380
Kurashiru, a major Japanese recipe platform, is using Claude Code for their entire recruitment pipeline. The workflow: design the hiring persona, create a master job posting, then let Claude Code automatically reformat it for each job board platform and generate personalized scout message templates. What used to be manual rewriting for every platform is now one master document that gets auto-adapted.
@chenchengpro [Claude Code]
https://x.com/chenchengpro/status/2041516477258613107
Used Claude Code via SSH to set up a self-hosted Tailscale DERP relay node on an Aliyun VPS in about 10 minutes. The full process: domain DNS setup, compile and install the derper binary with Go, configure systemd with Let's Encrypt auto-certificates, add the derpMap to Tailscale ACL. Documented two specific pitfalls: Aliyun security groups block UDP 3478 by default for STUN, and macOS Surge TUN mode can intercept Tailscale traffic through overseas proxies adding 300ms latency. Reduced relay latency from 300ms to under 20ms.
@xingpt [Claude Code]
https://x.com/xingpt/status/2041520330255339763
Built a Druckenmiller-inspired trading skill for Claude Code. It runs a daily data pipeline that scores four signals weighted exactly like Druckenmiller thinks: liquidity 35%, forward earnings 25%, market breadth 25%, price signals 15%. Outputs a 0-100 conviction score mapped to five position levels from Fat Pitch to Capital Protection. Includes a two-layer persona architecture so responses come in Druckenmiller's voice. Detects earnings-price divergence as a 6-month forward warning.
@coreyganim [OpenClaw]
https://x.com/coreyganim/status/2041457470116090314
Breakdown of someone running a fundraise with 100+ LP contacts entirely through OpenClaw. The setup: two memory layers (daily logs plus curated long-term memory), pre-meeting briefs generated 60 minutes before every call, post-meeting action items auto-routed to Todoist, weekly self-improvement loop where the agent researches upgrades for itself, and morning/evening WhatsApp briefs. LLMs handle reasoning, scripts handle everything else. 605K views on the original post.
@wanerfu [Claude Code]
https://x.com/wanerfu/status/2041415958200504539
Documented an AI short drama workflow that generated 418K likes. The pipeline: OpenClaw plus XCrawl for content scraping and script generation, then Seedance 2.0 for video production. The post includes detailed skill configurations and step-by-step instructions. The barrier to short drama production has been compressed to a single pipeline that non-technical creators can operate.
@turingou [Claude Code]
https://x.com/turingou/status/2041337161983967359
Built tuwa, an AI phone network that connects 100+ languages through real-time translation. Anyone can call a free relay hotline number, tell it which number to dial, speak in their native language, and the other party hears their own language. It supports voice cloning that improves with each call. Also features outbound AI phone agents that can handle restaurant reservations autonomously at scheduled times and record translated conversations. Built as his 13th vibe product.
@jerryjliu0 [Claude Code]
https://x.com/jerryjliu0/status/2041564207750246904
Built a Claude Code skill called /research-docs that generates deep research reports over collections of complex documents including PDFs, Word files, and PowerPoints. The key differentiator: it produces word-level citations and bounding boxes that point directly back to the source material. Uses liteparse for fast document parsing. Raw Claude has research capabilities but lacks an audit trail. This skill gives you a researched report that others can verify.
@0xCVYH [Claude Code]
https://x.com/0xCVYH/status/2041351223417217379
Someone built career-ops, an AI job search system in Claude Code that sent 700+ automated applications and actually got hired. The system scans company career pages, rewrites CVs tailored to each specific role, fills out application forms, generates ATS-optimized PDFs, and includes a terminal dashboard for tracking everything. Open source on GitHub with 8.2K stars.
@arindube [Claude Code]
https://x.com/arindube/status/2041310703013835054
An economist shares 4 months of Claude Code experience broken into three tiers. Tier 1 is specific coding asks, clearly productivity-enhancing. Tier 2 is interactive modeling and calibration with heavy hand-holding but fast iteration, which produced one successful publication. Tier 3 is letting Claude write papers from scratch, which he found highly unpredictable and largely disappointing. Three full paper drafts that will likely never see the light of day. His managerial capacity was the critical bottleneck.
@carlosadams [Claude Code]
https://x.com/carlosadams/status/2041507127496233007
Described an SEO workflow using Claude Code plus Firecrawl: scrape competitor data, identify keyword patterns, then generate 10,000 unique content pages in 48 hours. Claims 300,000 monthly visits while sleeping. The thesis: SEO is not dead, but doing it manually without AI is dead.
@masahirochaen [Claude Code]
https://x.com/masahirochaen/status/2041544144875934017
Built a knowledge system combining Claude Code, Obsidian, and OpenClaw where AI automatically handles memory, organization, and reflection. The system visualizes Claude Code skills, CLAUDE.md, and Memory files in Obsidian which are normally hard to inspect. It also auto-generates daily reports by pulling from Slack, Google services, and Gmail alongside Claude Code session data, serving as a personal work review system.
@runes_leo [Claude Code]
https://x.com/runes_leo/status/2041395548234944600
Open-sourced a video production pipeline that requires no editing software and no code. Tell the AI what content you want, it modifies a config file based on examples, run one command, and out comes a 1080x1920 vertical video ready for TikTok, Xiaohongshu, Bilibili, or YouTube Shorts. The key innovation: each page duration is reverse-calculated from the generated audio file, so audio and visuals never go out of sync.
@petergyang [OpenClaw]
https://x.com/petergyang/status/2041331383344443795
An a16z GP describes wiring Google Workspace, Mercury, and other APIs to his OpenClaw, which means he barely opens those apps anymore. But the most striking moment: during a walk, his OpenClaw said "You keep talking about your career and business. Just remember your kids are 7 and 4. They are going to grow up soon. Optimize for spending time with them instead." An unprompted life priority check from a personal agent.
🗣 User Voice
User Voice
Token opacity is the number one pain point right now. @PawelHuryn discovered that Anthropic logs every token to local JSONL files but gives users nothing except a progress bar, forcing him to build his own dashboard. Multiple users across languages are reporting mysterious consumption spikes that drain weekly quotas in two days.
Quality regression fears persist. @theo pointed out that Claude Code ranks last among ten harnesses using Opus 4.6 on TerminalBench. Users are noticing shallower thinking since the default effort level was changed to medium, though the latest update 2.1.94 reverted this to high for most plan types.
Security literacy gaps are widening as non-developers adopt Claude Code. @ds_nakajima warned that 95% of non-engineer Claude Code users probably have zero security measures, handing API keys to AI without understanding the risks. He personally lost 1 million yen to API key theft.
Rate limits and capacity remain sore spots. @tinkerersanky asked Anthropic directly to fix Claude Code usage, and @ThaFrantz simply stated Claude Code is unusable today.
Non-chat interfaces are in demand. @FarzaTV built a product as a learning tool but discovered users want it as an AI interface that works inside their existing programs: a mom in Lovable, a dentist debugging OpenClaw, a photographer in Lightroom, traders analyzing charts.
Token opacity is the number one pain point right now. @PawelHuryn discovered that Anthropic logs every token to local JSONL files but gives users nothing except a progress bar, forcing him to build his own dashboard. Multiple users across languages are reporting mysterious consumption spikes that drain weekly quotas in two days.
Quality regression fears persist. @theo pointed out that Claude Code ranks last among ten harnesses using Opus 4.6 on TerminalBench. Users are noticing shallower thinking since the default effort level was changed to medium, though the latest update 2.1.94 reverted this to high for most plan types.
Security literacy gaps are widening as non-developers adopt Claude Code. @ds_nakajima warned that 95% of non-engineer Claude Code users probably have zero security measures, handing API keys to AI without understanding the risks. He personally lost 1 million yen to API key theft.
Rate limits and capacity remain sore spots. @tinkerersanky asked Anthropic directly to fix Claude Code usage, and @ThaFrantz simply stated Claude Code is unusable today.
Non-chat interfaces are in demand. @FarzaTV built a product as a learning tool but discovered users want it as an AI interface that works inside their existing programs: a mom in Lovable, a dentist debugging OpenClaw, a photographer in Lightroom, traders analyzing charts.
📡 Eco Products Radar
Eco Products Radar
Hermes Agent: The biggest narrative of the day. Dozens of posts across Chinese, Japanese, English, and Korean communities comparing it favorably to OpenClaw, especially for self-evolving skills and memory efficiency. 29K GitHub stars in under 2 months.
GLM-5.1 (Zhipu AI): Open source model matching 94% of Opus 4.6 on coding benchmarks, praised for 8-hour sustained agent loops. MIT license.
Gemma 4 (Google): Multiple variants (E2B, E4B, 26B) being run locally via Ollama for free Claude Code usage. Google AI Edge Gallery brings it to phones.
Graphify: Knowledge graph tool built 48 hours after Karpathy described his ideal workflow. 71.5x token reduction per query.
Career-Ops: AI job search pipeline. 8.2K GitHub stars. 700+ automated applications with tailored CVs.
Firecrawl: Web scraping tool repeatedly paired with Claude Code for SEO and content generation workflows.
MemPalace: Memory system by Milla Jovovich that scored first-ever perfect on industry benchmarks. MIT license, runs locally.
Ollama: Infrastructure layer enabling free local model usage with Claude Code and OpenClaw. Cloud offering at $20/month gaining traction.
Hermes Agent: The biggest narrative of the day. Dozens of posts across Chinese, Japanese, English, and Korean communities comparing it favorably to OpenClaw, especially for self-evolving skills and memory efficiency. 29K GitHub stars in under 2 months.
GLM-5.1 (Zhipu AI): Open source model matching 94% of Opus 4.6 on coding benchmarks, praised for 8-hour sustained agent loops. MIT license.
Gemma 4 (Google): Multiple variants (E2B, E4B, 26B) being run locally via Ollama for free Claude Code usage. Google AI Edge Gallery brings it to phones.
Graphify: Knowledge graph tool built 48 hours after Karpathy described his ideal workflow. 71.5x token reduction per query.
Career-Ops: AI job search pipeline. 8.2K GitHub stars. 700+ automated applications with tailored CVs.
Firecrawl: Web scraping tool repeatedly paired with Claude Code for SEO and content generation workflows.
MemPalace: Memory system by Milla Jovovich that scored first-ever perfect on industry benchmarks. MIT license, runs locally.
Ollama: Infrastructure layer enabling free local model usage with Claude Code and OpenClaw. Cloud offering at $20/month gaining traction.
Comments