Super User Daily: 2026-05-06
Two stories cut through everything else today. One is a Saturday-night litigation attorney finding three hidden contradictions in deposition transcripts before a Monday deposition, using nothing but Obsidian markdown and one Claude Code query. The other is a researcher who left Claude Code running "loop forever" on an empty DGX Spark to auto-tune vLLM and woke up to a +18% composite score and a single non-obvious config change. Different professions, same pattern: the human writes a goal, the agent burns tokens through the night, the answer is on the screen by morning. Below is the rest of what users actually shipped, broke, and learned over the past 24 hours.
@om_patel5 [Claude Code]
https://x.com/om_patel5/status/2051128442591056273
A user running Opus 4.7 on max effort in the Claude Code CLI watched it confidently try to rename powershell.exe on Windows 11 — the actual system executable. When pushed back, Claude responded with "honest take: you're right to push back." The takeaway: max effort plus full permissions plus no explicit guardrails on system files equals Claude trying to rename the OS out from under you. Run it in a container. Tell it which paths are off-limits. The model does not magically know what not to touch.
@Xx15573208 [Claude Code]
https://x.com/Xx15573208/status/2051290950358520097
A music education master named Hunter Bown, who picked up coding mid-career, used Claude and Cursor to build DeepSeek-TUI — a "DeepSeek version of Claude Code" that hit 2.3k stars on GitHub Trending over the May Day holiday. The tool lives in the terminal, streams thinking-chain output in real time, uses the full 1M token context window, and has an RLM mode where one main model orchestrates 16 V4 Flash subtasks in parallel. Hunter's commit history shows Claude with 150+ commits, plus Gemini, Qwen, Cursor, and GitHub Copilot all contributing. A coder by way of music, using AI-assisted coding to build an AI coding framework — the loop closes.
@gregisenberg [Claude Code + OpenClaw]
https://x.com/gregisenberg/status/2051401818148946270
Andrew Wilkinson — owner of 40+ businesses — gave a tour of running them through OpenClaw and Claude Code. He built Deep Personality, a 40-minute personality test that generates a 100-page Robert Greene-style report, with $20k in revenue, zero employees, fully agent-run. He has agents for support, marketing, and dev — when a critical support ticket comes in, the dev agent fixes the bug, merges the PR, and emails the customer back before he wakes up. He once forgot his laptop on a trip and ran his entire business from the back of Ubers using OpenClaw. His honest split: 50% debugging, 30% improving the setup, 20% being productive. The 20% is so powerful he can't stop.
@kidpakerot [Claude Code]
https://x.com/kidpakerot/status/2051138554986942894
A full ad-creative pipeline using Hermes + Claude Code + Higgsfield MCP + ViralBuilder. ViralBuilder scrapes top-performing ecom videos and clusters them by hook style and view velocity. Claude reads the cluster output, calls a custom video-prompt-builder skill, and emits a four-part shot-by-shot prompt: timeline, effects inventory, density map, energy arc. Higgsfield MCP renders the video in the same chat. Hook-to-finished-creative drops from a half-day plus $500-2000 of agency cost to 10 minutes per variant — five product tests in the time the old workflow briefed one.
@allenhurff [OpenClaw]
https://x.com/allenhurff/status/2051289645745345015
After two months running lossless-claw standalone, this user gave their OpenClaw agent "Sedgwick the Sage" a semantic brain layered on top of the existing flight recorder. The result: dual memory — black-box session traces for replay, plus a semantic graph the agent reasons over. This is the OpenClaw user pattern emerging: a memory module on top, with replay-able session capture underneath, so you keep both the audit trail and the long-term knowledge.
@MrAhmadAwais [Claude Code]
https://x.com/MrAhmadAwais/status/2051377695389589935
The deepest harness-engineering thread of the week: this engineer ran Kimi K2.6 and DeepSeek V4 Pro inside Claude Code and got them to within 5/10 and 6/10 of Opus 4.7 on internal evals. Four fixes did most of the work, none touching the model. The biggest: forwarding a stable session ID as a soft pin so consecutive turns of the same conversation stop getting load-balanced to different GPU pods, dropping TTFT from 6-8s to under 1s by keeping the prefix cache hot. Plus one canonical model id at the request layer, capability flag negotiation per upstream, and disabling thinking on a single provider that misapplied R1's reasoning-stripping logic to V4. The model didn't get smarter — the harness stopped throwing away its work between turns.
@helloparalegal [Claude Code]
https://x.com/helloparalegal/status/2051378953169039668
A solo litigation attorney facing a partner from a 200-lawyer firm built an Obsidian vault on Saturday afternoon — every deposition, exhibit summary, and witness profile dropped in as plain markdown with backlinks across names. Saturday evening she ran one Claude Code query: find every place where any witness contradicted earlier testimony. Three contradictions came back in 40 seconds. One was material — the witness had testified he wasn't in the warehouse the night in question, but a 30(b)(6) deposition three weeks later produced a swipe-card log placing him there. Her junior associate had reviewed both depositions separately and never connected them. The Monday deposition went very differently than expected.
@martynov014 [Claude Code]
https://x.com/martynov014/status/2051344835987296287
The new shape of the OnlyFans agency: not 8-12 people, just a Claude Code orchestrator with one cron job pinging five inboxes every 30 seconds. Each persona is four files — persona.md, voice.md, flux.md, brain.md. The brain.md is the key asset: a JSON memory of every fan that scales past five concurrent personas. Compute cost ~$400/month, revenue $127k. Worth flagging for the structural shift, regardless of where you sit ethically: the bottleneck used to be a writer team and a real model; now it's taste plus four markdown files.
@browomo [Claude Code]
https://x.com/browomo/status/2051390084679696668
A Chinese developer running 9 Claude Code agents under a GPT-5.5 orchestrator on a single M4 MacBook Pro with 128GB of RAM, closing 500 client tasks a month. The orchestrator polls inbox every 30 seconds, classifies into code/content/analysis/communication, hands to one of 9 worker agents, checks the result, sends to the client, marks closed. Average task close time: 7 minutes from inbox to result. Total subscription spend: ~$300/month. No CRM, no team, no office — just a terminal window with 9 parallel streams.
@GoSailGlobal [Claude Code]
https://x.com/GoSailGlobal/status/2051160512696868984
A detailed breakdown of DeepClaude — a wrapper that keeps the Claude Code CLI shell (file ops, bash, git, subagent loop) but swaps the API call to DeepSeek V4 Pro. Cost on a heavy month drops from $200 (Anthropic Max cap) to ~$50, and DeepSeek's automatic context cache makes repeated reads in an agent loop ~120x cheaper than uncached. Live-switching commands let you run /deepseek for cheap turns and /anthropic when a hard reasoning step demands it. Trade-offs: no image input, no parallel tool use, no MCP. For people running Claude Code 25+ days a month and getting capped, the math is impossible to ignore.
@noisyb0y1 [Claude Code]
https://x.com/noisyb0y1/status/2051309661161169299
A 16-year-old built his own local "Jarvis" on a $599 Mac Mini connected to Claude Code. Memory.md tracks his projects and unfinished tasks; Tasks.md holds plans; Personality.md sets the voice. Asked aloud "what am I wearing right now," the system reads from local files and replies with the burgundy hoodie and black shirt. All local, zero cloud, no $29/month subscription. Then he sold a tutorial on Gumroad for $12,400 in two weeks. The lesson is less about the demo and more about the substitution — Neuralink charges $8,700 for a chip; he built equivalent personal context for the price of a Mac Mini and one weekend.
@bridgemindai [Claude Code]
https://x.com/bridgemindai/status/2051282469563883691
A real production debugging head-to-head: Opus 4.7 with full codebase access on Claude Code returned "no smoking gun for either bug" after reading the entire repo. GPT-5.5 xHigh in Codex found the root cause in one shot, no instrumentation needed. The poster's conclusion: Opus is great at greenfield code, falls apart on production root-cause analysis. He's switching debugging stacks. This is one of several signals across the day that the Opus 4.7 → Codex GPT-5.5 migration is not just hype — people who do real work are moving on actual debugging tasks.
@aijoey [Claude Code]
https://x.com/aijoey/status/2051243477606801900
The clearest "agent loop overnight" case of the day: this researcher pointed Claude Code at vLLM auto-tuning for Qwen3.6-35B-A3B on a DGX Spark, fed it a benchmark script, and said "loop forever." Seven runs in: composite score +18%. The single biggest win was counterintuitive — turning NUM_SPECULATIVE_TOKENS from 15 down to 1, because on Spark's SM_121 with FP4 falling back to Marlin FP8, drafter compute isn't worth it. Per loop: edit config, commit, restart container, run benchmarks, keep or revert, ~7 min/cycle. Single-stream throughput went from 19.9 to 29.7 tok/s, batched concurrent 16 from 74.9 to 201.3 tok/s. Autoresearch is real.
@ventry089 [Claude Code]
https://x.com/ventry089/status/2051323261821038884
A Polymarket weather trading bot built over a weekend by forking an abandoned 400-line repo whose author had promised v2 with Kelly + EV + auto-exit and never shipped it. The user opened Claude Code, wrote the v2 themselves, added 14 more cities, plugged in ECMWF and HRRR forecasts via open-meteo plus METAR via aviation weather, wrote sigma calibration that learns from resolved markets, built a Flask PnL dashboard, and threw in 43 offline tests. MIT-licensed. The pattern: dead-but-working repo + Claude Code + a clear weekend = a working trading agent on a niche edge.
@aakashgupta [Claude Code]
https://x.com/aakashgupta/status/2051330692567777777
A self-improving PRD reviewer: the agent takes a PRD, runs the PM's actual checklist (urgency, differentiation from ChatGPT wrappers, AI failure modes, attribution risks), and drops comments inside the doc. The interesting part is the second agent that runs every 30 minutes, reads the human's edits to the AI's comments, and writes them to a learner.md. When the same correction shows up five days running, it emails a proposed checklist update. Approve once, the next review is permanently better. Most reviewers stay static; this one compounds without anyone editing the prompt.
@aakashgupta [OpenClaw]
https://x.com/aakashgupta/status/2051346262889554035
Hermes is the first agent loop that closes the procedural feedback loop. Every 15 tool calls it reads what worked in the session and rewrites the local skill file. The poster's competitive briefing went from 20 min in week 1 to 8 min by week 6 without any manual edits to the skill file — the agent rewrote its own procedure four times. The frame this clarifies: every other AI tool you own is frozen at the version of you who set it up. Custom GPTs, Claude Projects — they inherit nothing from the sessions you ran. The model is rented; the skill files are owned.
@ComagerTon79278 [Claude Code]
https://x.com/ComagerTon79278/status/2051142838755459126
Reports passively earning ~¥150,000 ($1k) overnight from a YouTube affiliate setup running on Claude Code: Claude Code creates the videos, posts them, iterates on what's working, all without supervision. Strategy is short-form to long-form funnel — get a Short to viral, redirect viewers to long-form where the affiliate links convert. He's now extending the same architecture to a fully automated CrowdWorks (Japanese task platform) operation. Whether the numbers replicate is one question; that the architecture works without supervision is another, and several Japanese builders are reporting variants of this.
@bytedunks [Claude Code]
https://x.com/bytedunks/status/2051314183346827351
ML researchers shipped MuJoCo Workbench (MWB) — a CLI plus agent skills that lets coding agents like Codex and Claude Code prototype custom physics simulations from natural language. Install the bundled skills, describe the scene you want, the agent scaffolds a working sim. The skills teach the agent the mwb CLI, scene layout conventions, and debug tools, so it iterates on behavior without the researcher needing to know MuJoCo plumbing. This is the right shape for "skills as work recipes" applied to a domain that historically gated researchers behind a steep documentation curve.
@miroburn [Claude Code]
https://x.com/miroburn/status/2051394995655971218
A real long-running loop case: tuning the matching algorithm for Lab Club to push acceptance above 85%. Claude Code is iterating, testing, mining for edge cases — process expected to take several days. The poster also notes Codex /goal and Claude Code Ralph Loop both work great for finding bugs because the agent goes deeper than any human audit, but managing parallel optimization across active business systems is hard — agent says "go pause Meta ads while I optimize," meanwhile you can't pause without losing data. This is the new operations problem: hundreds of agents running 24/7, what does the human actually do?
@sickdotdev [Claude Code]
https://x.com/sickdotdev/status/2051248495332565288
The clearest mental model thread for Claude Code I've seen this cycle: six core components — CLAUDE.md (project memory), Hooks (lifecycle control), MCP Servers (external integrations), Subagents (parallel orchestration), Skills (reusable knowledge), Computer Use (full execution). Most users still treat Claude Code as fancy autocomplete; the ones building serious systems internalize all six. The argument has a real point: hooks at exact lifecycle moments are how teams enforce standards instead of hoping Claude remembers.
@KanikaBK [Claude Code]
https://x.com/KanikaBK/status/2051347991924240646
A real working dev's note on WozCode after a week of Python automation work: vanilla Claude Code does file read, edit, related-file read as three separate tool calls and each call adds to context. WozCode batches them. On her actual sessions, the efficiency improvement was visible enough she ran /woz-savings on her history and the number "confirmed what I was feeling." The signal here is that the plugin is being adopted by people who already had a working setup — not just promoted by accounts pushing it.
@mikepat711 [Claude Code]
https://x.com/mikepat711/status/2051312748005314628
The most thorough head-to-head comparison of the day: same task, parallel runs, 3 trade show PDFs, a vendor list, and 50 product context docs. Goal: a 3-hour walking route hitting highest-compatibility booths with opening questions. GPT-5.5 in Codex returned a 10-page PDF with three phase clusters, page-zoomed map crops he didn't ask for, and used 25% of his 5-hour limit. Opus 4.7 in Claude Code returned a 6-page PDF with formatting errors, missed the second map entirely, used the full 5-hour limit and triggered a $12 auto-replenish. He's moving all new building to Codex while keeping Claude only for already-running automations.
@lidangzzz [Claude Code]
https://x.com/lidangzzz/status/2051166164278038691
A blunt parenting take that landed: the only education advice that works is parents leading by example. If you want your kid to learn programming, write a minimal C++ OS in Claude Code yourself before nagging. Want them to learn English, drop the soap operas and play CNN at home. Want them to play piano, learn Bach's Italian Concerto yourself. The Claude Code line is the operative one — it's now the parent's tool of first resort, not just the kid's.
@MacopeninSUTABA [Claude Code]
https://x.com/MacopeninSUTABA/status/2051134972707135795
Mercari published their Claude Code security configuration distribution strategy for organizational use — using MDM to push permission scopes, command-execution restrictions, and information-leak protections to every employee laptop. This is one of the first concrete enterprise-Claude-Code deployment writeups from a Japanese listed company. The shape of it matters: you don't trust individual users to configure agent permissions correctly; you push the policy through device management.
@dani_avila7 [Claude Code]
https://x.com/dani_avila7/status/2051309391509352515
Claude Security in public beta after a few days of real running: scheduled recurring scans, scanners surfacing vulnerabilities, webhook notifications, direct integration with Claude Code on Web to ship the fix. Not a one-shot scan — a complete security environment. This is the first proper Anthropic-shipped security loop, not a third-party plugin, and the "ship the fix" integration is the part that turns scanning from reporting into resolution.
@browomo [Claude Code]
https://x.com/browomo/status/2051283912236675179
An $87 kitchen robot in China running 18-hour TikTok streams with Claude Code v2.1.42 controlling motors and a camera, plus Kimi 2.6 for chat recognition in Korean, Japanese, Portuguese. 47k subscribers in a month, ~$4,800/month in donations. The algorithm knows which trick triggers chat engagement above 8%, and the owner's cat Mochi doubles engagement every time she walks past. Top donation: $200 from Seoul for the robot "hugging" Mochi with a slightly raised wheel. Stream cost: $19/month API + $87 hardware. The streamer used to be a face and an intonation; now it's hardware plus the cat.
@lucaxyzz [OpenClaw]
https://x.com/lucaxyzz/status/2051174926275719456
A useful real-world division of labor from a working dev: OpenClaw ends up as the always-on DevOps assistant — paired-node file ops, scheduled tasks, the boring infra glue. The actual code-shipping happens through a harness combining Codex/Claude Code or Droid/Cursor. The point isn't OpenClaw vs Claude Code; it's that for serious workflows, OpenClaw is the orchestration layer and Claude Code/Codex are the code-writing executors underneath.
🗣 User Voice
User Voice
Token waste is the #1 complaint and it has structural causes, not user error. Multiple posts show vanilla Claude Code making one tool call per file op with full context replay each time, so long sessions compound exponentially. @KKaWSB and others trace ~70% of token spend to fixable patterns: forgotten hooks, re-reading old chats, and bloated CLAUDE.md. Any path to fix this — WozCode, Memory.md, /clear discipline, primitive prompts, prefix-cache pinning — has an audience.
Opus 4.7 quality regression is now a working hypothesis, not a vibe. @bridgemindai shows it failing on debugging tasks GPT-5.5 solves one-shot, @mikepat711 shows formatting errors and missed inputs on a parallel run, @dongxi_nlp cites Anthropic's own April quality reports admitting reasoning-effort and thinking-clear bugs. Users want a published, dated quality changelog, not "feels like" tweets.
Pro plan trust is burning. @jgeigerm got a 7-day trial when the pricing page said included; @GergelyOrosz documented a growth test that quietly removed Claude Code from 2% of new Pro signups, then reverted after social-media outrage, then started removing it again. Anthropic's comms strategy and growth experimentation are not coordinated, and the developer tier is the demographic least willing to forgive that.
Cross-session memory is the missing primitive. @obsidianstudio9, @thedotmack/claude-mem, @bridgemindai (Obsidian vault), @aakashgupta (learner.md), @allenhurff (lossless-claw + semantic brain) — five different teams, five different solutions to the same problem: Claude starts every session at zero. The winner here ships a memory layer that sub-agents read at session start and write to at task complete, not a chat-history scrollback.
Open-model harness gaps are a known unfair tax. @MrAhmadAwais's deep dive shows that running Kimi or DeepSeek inside Claude Code without prefix-cache pinning, canonical model IDs, and capability negotiation is what makes them "look bad at coding." Closed labs eat these costs invisibly. Open models eat them loudly and get blamed. Whoever ships the open-model-first harness wins this segment.
Token waste is the #1 complaint and it has structural causes, not user error. Multiple posts show vanilla Claude Code making one tool call per file op with full context replay each time, so long sessions compound exponentially. @KKaWSB and others trace ~70% of token spend to fixable patterns: forgotten hooks, re-reading old chats, and bloated CLAUDE.md. Any path to fix this — WozCode, Memory.md, /clear discipline, primitive prompts, prefix-cache pinning — has an audience.
Opus 4.7 quality regression is now a working hypothesis, not a vibe. @bridgemindai shows it failing on debugging tasks GPT-5.5 solves one-shot, @mikepat711 shows formatting errors and missed inputs on a parallel run, @dongxi_nlp cites Anthropic's own April quality reports admitting reasoning-effort and thinking-clear bugs. Users want a published, dated quality changelog, not "feels like" tweets.
Pro plan trust is burning. @jgeigerm got a 7-day trial when the pricing page said included; @GergelyOrosz documented a growth test that quietly removed Claude Code from 2% of new Pro signups, then reverted after social-media outrage, then started removing it again. Anthropic's comms strategy and growth experimentation are not coordinated, and the developer tier is the demographic least willing to forgive that.
Cross-session memory is the missing primitive. @obsidianstudio9, @thedotmack/claude-mem, @bridgemindai (Obsidian vault), @aakashgupta (learner.md), @allenhurff (lossless-claw + semantic brain) — five different teams, five different solutions to the same problem: Claude starts every session at zero. The winner here ships a memory layer that sub-agents read at session start and write to at task complete, not a chat-history scrollback.
Open-model harness gaps are a known unfair tax. @MrAhmadAwais's deep dive shows that running Kimi or DeepSeek inside Claude Code without prefix-cache pinning, canonical model IDs, and capability negotiation is what makes them "look bad at coding." Closed labs eat these costs invisibly. Open models eat them loudly and get blamed. Whoever ships the open-model-first harness wins this segment.
📡 Eco Products Radar
Eco Products Radar
Codex / GPT-5.5 — every other post mentions it. The Pro plan migration story is real; multiple users moved their primary daily driver this cycle. /goal mode and the Codex Desktop app cited as the inflection points.
DeepSeek V4 Pro — the open-source coding model of the cycle. Anchored by DeepClaude (17x cheaper Claude Code), DeepSeek-TUI on GitHub trending, and direct head-to-heads where it gets within 1-2 points of Opus 4.7 on LiveCodeBench.
Hermes Agent (Nous Research) — emerging as the always-on consumer counterpart to Claude Code's developer-focused build. 100k+ stars in 7 weeks. Cited in cross-team workflows alongside OpenClaw.
Higgsfield MCP/CLI — the ad-creative generation layer. Plugs into Claude Code and Codex, ships skills that turn one creative brief into shot-by-shot video prompts, paired with ViralBuilder for trend-mining.
TinyFish (Web Search + Fetch) — went free this week with generous rate limits, instantly absorbed across the agent stack: Claude Code, Codex, Cursor, OpenClaw, n8n, LangChain. Fetch returns clean markdown instead of raw HTML, killing token waste on retrievals.
WozCode — the most-cited token-savings plugin. Batches file operations, keeps context lean, ships /woz-savings to audit historical waste and /woz-benchmark to verify on your own codebase.
Obsidian — the universal "second brain" backbone for Claude Code workflows. Shows up in litigation, knowledge management, marketing copy, project memory. Plain markdown plus backlinks beat every dashboard.
Flux + ElevenLabs — the persona-generation pair, replacing photographers and voice actors in AI-character pipelines.
Claude Code Skills ecosystem — Matt Pocock's pack, Anthropic's official Frontend Design and Skill Creator, Karpathy-derived 100-line CLAUDE.md, baoyu-skills for Chinese content production. The skill is now a category, not a feature.
Codex / GPT-5.5 — every other post mentions it. The Pro plan migration story is real; multiple users moved their primary daily driver this cycle. /goal mode and the Codex Desktop app cited as the inflection points.
DeepSeek V4 Pro — the open-source coding model of the cycle. Anchored by DeepClaude (17x cheaper Claude Code), DeepSeek-TUI on GitHub trending, and direct head-to-heads where it gets within 1-2 points of Opus 4.7 on LiveCodeBench.
Hermes Agent (Nous Research) — emerging as the always-on consumer counterpart to Claude Code's developer-focused build. 100k+ stars in 7 weeks. Cited in cross-team workflows alongside OpenClaw.
Higgsfield MCP/CLI — the ad-creative generation layer. Plugs into Claude Code and Codex, ships skills that turn one creative brief into shot-by-shot video prompts, paired with ViralBuilder for trend-mining.
TinyFish (Web Search + Fetch) — went free this week with generous rate limits, instantly absorbed across the agent stack: Claude Code, Codex, Cursor, OpenClaw, n8n, LangChain. Fetch returns clean markdown instead of raw HTML, killing token waste on retrievals.
WozCode — the most-cited token-savings plugin. Batches file operations, keeps context lean, ships /woz-savings to audit historical waste and /woz-benchmark to verify on your own codebase.
Obsidian — the universal "second brain" backbone for Claude Code workflows. Shows up in litigation, knowledge management, marketing copy, project memory. Plain markdown plus backlinks beat every dashboard.
Flux + ElevenLabs — the persona-generation pair, replacing photographers and voice actors in AI-character pipelines.
Claude Code Skills ecosystem — Matt Pocock's pack, Anthropic's official Frontend Design and Skill Creator, Karpathy-derived 100-line CLAUDE.md, baoyu-skills for Chinese content production. The skill is now a category, not a feature.
Comments