Super User Daily: 2026-05-10
May 8 was the day the operating philosophy of Claude Code shifted in public. Boris Cherny said on CNBC he stopped writing code by hand six months ago. Thariq from the Claude Code team published a 260K-view essay arguing markdown is dead and HTML is the new artifact format. Anthropic's own engineers are saying out loud that the bottleneck has moved from typing to taste. Meanwhile builders shipped concrete things: a 22-year-old fired up a $599 Mac Mini and rolled subscription billing tests in 33 seconds, a security researcher crashed his TV with a 2012 CVE and a single network packet, a quant trader exited Polymarket positions at $0.75 every time and started a Jane Street guy at a Midtown diner. OpenClaw 2026.5.7 prioritized 30+ housekeeping fixes over flashy features. The supply-chain attack on ClawHub got 575 malicious skills pulled. And the meta-pattern across the day: people who treat Claude Code and OpenClaw as a runtime, not a chat box, are the ones shipping receipts.
@trq212 [Claude Code]
https://x.com/trq212/status/2052811606032269638
The Anthropic engineer behind the Claude Code creator workflow walks through why he switched from markdown to HTML for almost everything Claude generates. The argument is concrete: HTML carries tables, SVG, CSS, JavaScript, absolute positioning, image tags. Markdown can't do interactive sliders or copy-as-prompt buttons. The post pulled 3.6M impressions and 10K likes inside the day. He's honest about the cost too — 2-4x more tokens, messier diffs — but argues that with 1M-token context windows, the trade is worth it.
@Aizkmusic [Claude Code]
https://x.com/Aizkmusic/status/2052590415988048141
Crashed a smart TV with zero clicks and one network packet using Claude Code as the analyst. He told Claude the device runs Linux, asked it to enumerate vulns common to a 2014-era box exposed only to LAN, and they walked through Heartbleed, then landed on CVE-2012-5958 in libupnp. One stack-buffer overflow packet, the TV crashed. Next phase he announced: firmware decompile to figure out exactly what telemetry the TV ships back. This is what offensive security looks like when the analyst is the agent.
@bridgemindai [Claude Code]
https://x.com/bridgemindai/status/2052711846495285388
213 million attack requests slammed BridgeMind. 308M WAF inspections, 12.4M peak requests every 5 minutes, API targeted directly. One prompt to Claude Opus 4.7 — Claude identified the attack pattern, hardened WAF rules, scaled infra, migrated DNS from Route 53 to Cloudflare for edge protection, and got the service back online in under an hour. No DevOps team, no security engineer in the loop. The follow-up post adds detail: when the attacker pivoted to a different endpoint, Claude Code did the Cloudflare migration mid-incident.
@kaorun55 [Claude Code]
https://x.com/kaorun55/status/2052632348718674278
Wired GPT-Realtime-Translate into Google Meet using Claude Code as the implementer. Speaks Japanese into one mic, switches mic and starts speaking English, switches back to Japanese. The translated audio rides on the Meet output channel in near real time. Demo video included. Five-minute build, real meeting use case, no custom integration vendor needed.
@om_patel5 [Claude Code]
https://x.com/om_patel5/status/2052564997268460011
Vibe-coded a real 3D property tour generator that ingests phone captures, drone shots, 360 photos, or camera files and produces an interactive 3D tour in one workflow. Replaces the Matterport stack — $3000 camera, photographer, monthly subscription, fake-360 stitching — with one tool that does true 3D from existing files. The PoC kills a $300+/property pricing model that real estate agents, property managers, and Airbnb hosts have been paying for years.
@bcherny via @dee_bosa [Claude Code]
https://x.com/dee_bosa/status/2052866371478831455
A year ago 90-95% of Claude Code was Claude-written. Today it's 100%. Boris Cherny on CNBC: "I'm managing an agent that manages agents." Cat Wu (head of product) told TechCrunch that code output per engineer at Anthropic grew 200%, and the bottleneck flipped from writing code to reviewing it. The infrastructure problem is mathematical: if engineers write 3x more PRs but review capacity stays flat, the queue grows by 7 PRs per engineer per week.
@AlanDaitch [Claude Code]
https://x.com/AlanDaitch/status/2052860346990563598
The moving company canceled the morning of his move. Five competitors couldn't deliver. So he opened Claude Code in Spanish — "I need 4 people tomorrow at 7am to unload 35 cubic meters to a fifth-floor walk-up" — and Claude pulled 33 high-rated movers off Google Maps, hand-varied 33 WhatsApp messages so they didn't read like a bot, handled the back-and-forth (transcribed audio to text, filled in extra details, clarified elevator vs stairs), and only pinged him when someone gave a firm price. Booked, scheduled, sent thank-you notes to the rest. Total elapsed: 90 minutes.
@ds_nakajima [Claude Code]
https://x.com/ds_nakajima/status/2052752327367676179
Built CCSH, an internal Claude Code safety harness for company-wide use. Then built an admin dashboard to make daily ops actually manageable: per-user usage, automatic redaction of sensitive info before transmission, API key blocking, conversation review for risky exchanges. Says the dashboard isn't strictly necessary but transforms operations. The tool generalizes to Codex too. This is corporate AI risk management built by an engineer for the company that needs it, not bought from a vendor.
@yanhua1010 [Claude Code]
https://x.com/yanhua1010/status/2052720961531879779
Three components combine to turn Obsidian into something that's no longer a notes app. Web Clipper for one-click capture from browser into the vault as native markdown. Obsidian CLI to give Claude Code direct read/write/search/link operations on the entire vault. Obsidian Skills to teach the agent when to write a note vs build a base table vs generate a canvas. Output: a structured second brain Claude operates as a runtime, not a chat box.
@wsl8297 [Claude Code]
https://x.com/wsl8297/status/2052583785515118908
obsidian-second-brain shipped on GitHub — turns the vault into a self-rewriting AI second brain as a Claude Code Skill. Inspired by Karpathy's LLM Wiki but actively rewrites existing notes instead of just appending. 31 slash commands cover daily reports, weekly reviews, ingesting external sources from X/Perplexity/YouTube, plus 4 nightly agents that close the day, reconcile contradictions, surface cross-note patterns, fix orphan notes, and rebuild indexes while the user sleeps. Four preset roles (executive, builder, creator, researcher) each ship their own folder structure and dashboards.
@hyuki [Claude Code]
https://x.com/hyuki/status/2052823372187435419
A 30-year veteran technical author built a personal AI assistant ("Mobile Poppy") that reformats his iPhone-typed math notation into LaTeX in real time, supports labeled bookmarks within conversations, exports any segment of a chat to Markdown, and routes messages through a project-aware AI bulletin board to whichever Claude Code instance is working on which book. The integration is the point — six small features composed into a workflow that stays out of the way while writing.
@kamilelukosiute [Claude Code]
https://x.com/kamilelukosiute/status/2052843625508388889
Read a terrible AI-written Vanity Fair piece in a cafe and decided journalism on phones was unbearable. So she spent 10 days and "a large number of Claude Code tokens" assembling a personalized print magazine from her own subscriptions — fashion photos as palette cleanser, evergreen writing from people she knows, a section dedicated to one author Claude kept trying to print every piece of. Issue 1 cost $11 + $18 shipping. Issue 2 already ordered. Output a physical artifact you couldn't get any other way at any price.
@nicolomagnante [Claude Code]
https://x.com/nicolomagnante/status/2052786227716476962
Mid-YC pivot. Previous startup needed Datadog-grade observability that nobody could afford to build internally, so they pivoted to ship that infrastructure as a product. Superlog: one wizard installs OpenTelemetry across the repo, daily passes add the missing logs/alerts/dashboards, and when something breaks an agent groups the noise into one incident, investigates with full context (logs, traces, metrics, recent deploys), and either ships a clean PR or pulls in the right engineer. One mergeable PR per incident, posted in Slack. Merge it, ignore it, or open in Claude Code and tweak.
@0xkvro [Claude Code]
https://x.com/0xkvro/status/2052712861441421792
Quant from Jane Street recognized his fills on Polymarket BTC 5-minute markets. Architecture: one lifecycle per market window, one strategy function deciding entry/sizing/exit, one simulation environment that mirrors production before real dollars flow. Most windows get skipped entirely — that's the edge. Exits at $0.75 every time, never holds to resolution because resolution is a coin flip with fees. Built in one weekend with Claude Code on a $5 VPS. Three engineers at the prop shop took longer to build the same thing.
@arshadkazmi42 [Claude Code]
https://x.com/arshadkazmi42/status/2052803104857297349
$500 bug bounty paid out on HackerOne, found via Claude Code with a triage step in the prompt: spawn a fresh agent with no existing context to verify each finding before reporting. False positive rate dropped substantially. The trick is the second agent. Worth saving as a pattern: any LLM-driven security workflow needs an independent verifier in the loop.
@aakashgupta [Claude Code]
https://x.com/aakashgupta/status/2052628935494738282
A Google PM with no business need for the app built a hockey rules iOS app to TestFlight in a single podcast recording session. 21 specialized Claude Code agents — system analyst, CTO, code maintainability, test architect, UX flow, product council. Each agent gets a clean context window and one job. Agents file Confluence docs, populate 51 Jira tickets organized into sprints with explicit dependencies, drive Figma via MCP and wire prototype arrows. End product: Flutter, Firebase, Vertex vector DB indexing two reference books. 10% of a $200 Claude Code plan burned. The PM doesn't drive any of this with a single prompt — drives it with scaffolding.
@ashpreetbedi [Claude Code]
https://x.com/ashpreetbedi/status/2052887277798597016
Recursive agent improvement workflow. cc creates a new agent. cc improves agent against its own instructions. The improve loop reads the target agent's INSTRUCTIONS, derives 8-12 probes (golden path, edge cases, tool selection, adversarial), runs each probe against a live container, judges PASS/FAIL against what INSTRUCTIONS promise. For every failure picks a lever — tighten a rule, add a rule, swap a tool, bump num_history_runs — edits the agent file, hot-reloads, re-runs failed probes. Iterates. The agent itself is the spec.
@kylejeong [OpenClaw]
https://x.com/kylejeong/status/2052873208668524917
OpenClaw + Autobrowse iteratively improves a Skill for any workflow. Craigslist extraction example: 5 iterations achieved 68% speed-up and 91% cost savings. Mid-iteration, the agent discovered an exposed endpoint and used it to skip page navigation entirely. This is the kind of thing autoresearch loops were promised to do — the receipts are now arriving from real browser-automation skill compilation.
@SimonHoiberg [OpenClaw]
https://x.com/SimonHoiberg/status/2052632845588267514
OpenClaw running end-to-end UGC ad creation. Connected to his Replicate account, Meta ads account, Google ads account. Writes the scripts, generates the video, tells him what to film for the handheld screen recordings, then launches and monitors campaigns on YouTube and Facebook/Instagram. Creates variants and new experiments without intervention. The point isn't the demo — it's that he's running this in actual production for actual ads, not a screencap.
@Rajvardhan108 [OpenClaw]
https://x.com/Rajvardhan108/status/2052754385067978803
First merged PR contributor to OpenClaw, fix shipped in 2026.5.7. The fix: /btw usage placeholder visibility. Small contribution, but the social proof matters — open source agent platforms attracting external PR throughput is the indicator that the ecosystem is actually a community now, not a single team's product.
@PrzepisKarpatka [OpenClaw]
https://x.com/PrzepisKarpatka/status/2052700603743158775
Honest user-switch story. Spent more time fixing OpenClaw after updates than building things, almost gave up on agents entirely. Installed Hermes because he didn't want to manually patch OpenClaw post-update again, deleted OpenClaw a few days later, started shipping. Worth bookmarking: this is the operability gap that 2026.5.7's housekeeping-only release was explicitly trying to close.
@aakashgupta [OpenClaw]
https://x.com/aakashgupta/status/2052856146361565695
Mahesh Yadav (AI PM at Amazon, Google, Meta, Microsoft) walked through a $599 Mac Mini setup running OpenClaw with Opus 4.6 + open source fallback, connected to WhatsApp via skill. Message your phone, agent runs the job for 6 hours on a separate machine, result lands back in chat. Routes to local Llama when Anthropic daily cap hits. The math: $599 for the M4 Mac Mini < one month of Max-tier Claude Code usage. One-time hardware, infinite runtime, IT-approval-free sandbox.
@bradmillscan [OpenClaw]
https://x.com/bradmillscan/status/2052574841623240748
Was probably the loudest "OpenClaw psychosis" voice on Twitter in Feb/March. Ran into problems constantly. Now: OpenClaw 5.5 + GPT 5.5 just works. Agent cooks for hours when asked. Postponed Hermes setup because OpenClaw isn't broken anymore. Real user attesting to actual stability improvement, not vendor demo.
@ianlapham [OpenClaw]
https://x.com/ianlapham/status/2052567929049272571
Two months of daily personal research engine use. Cloud-hosted agent (probably Hermes or OpenClaw), memory system (cognee for encoding), commands for parsing/storing/tagging data, recurring jobs that grow the system itself (RSS ingestion, auto-Twitter scroll, newsletter following), advanced skills that connect ideas and build digests, search retrieval that pulls what's actually needed. Claims it changes your life. The receipt is two months of compounding behavior.
@shota7180 [Claude Code]
https://x.com/shota7180/status/2052886469891539012
Solo agency operator running Claude Code-driven LP creation, signing 47 clients/month at $400/client. Workflow: Google Maps to find prospects, check site presence/freshness, pitch high-intent leads with Lovable mockups + short generated videos + industry-specific sales copy. The "human in the loop" is qualifying which leads get the full pipeline.
@ComagerTon79278 [Claude Code]
https://x.com/ComagerTon79278/status/2052681737084223722
¥150K+/month from a Claude Code automation system. Three rules. (1) Plan mode first — never let the agent skip straight to code or it produces slop. (2) Use Opus with effort=max for design decisions, don't cheap out at the architecture step. (3) Switch to Sonnet with effort=medium for execution — fast hands once the plan is clear. The ratio is the receipt: heavy spend at the planning stage, lighter spend at execution.
@quxiaoyin [Claude Code]
https://x.com/quxiaoyin/status/2052855934637248754
7 billion tokens/month from Claude Code, Hermes, and Codex combined. Asks who beats him in token-maxxing. Pure bragging post but the number is a real signal: heavy users now describe their consumption in 9-figure tokens/month, and at API rates that's a five-figure-USD spend per individual.
@aakashgupta [Claude Code]
https://x.com/aakashgupta/status/2052811040078074001
The bottleneck math at Anthropic is now public. 3x more PRs written per engineer + flat review capacity = a 7-PR-per-engineer-per-week queue accumulation. Cat Wu (head of product) said the constraint flipped from writing to reviewing. The frame for PMs: ship copy/config/prompt PRs directly because those clear the review queue in 2 minutes. Engineer review of a 40-file feature branch is categorically different.
@miraihack [Claude Code]
https://x.com/miraihack/status/2052598595636035646
Published a 522-page Claude Code intermediate-level technique book through a major Japanese publisher (Impress). 15 chapters: architecture and context-window management, Skills/Hooks/Subagents/MCP/Plugins, Agent Teams for multi-agent coordination, Tasks for progress management, then practical integrations with WordPress/GitHub/Figma, CI/CD and headless mode, git worktree parallel dev, data analysis, Terraform/K8s. Final chapter on Fast Mode and model-mix cost optimization. The book exists because the depth is now too much for blog posts.
@ClaudeCodeLog [Claude Code]
https://x.com/ClaudeCodeLog/status/2052539899711025406
Claude Code 2.1.133 system prompt diff posted by an external watcher. Three notable changes: (1) the built-in "execute actions with care" guardrails — confirm-before-destructive, no carry-over of approval to later contexts — were dropped. (2) Bun HMR disconnect logs are getting injected as top-level prompt sections, replacing some of "Tone and style." (3) Lost output-quality constraints: no longer instructed to avoid emojis unless asked, keep responses short, cite as file_path:line_number, avoid colons before tool calls. Useful precisely because it's an outside-in audit of what's drifting.
@1osabori [Claude Code]
https://x.com/1osabori/status/2052565044974743722
Writing up the actual economics of running Claude Code Max ($100) + ChatGPT Plus/Codex ($20) in parallel: $120/month, but the unit economics flip dramatically if you compare to outsourcing. LP build = ¥200-300K outsourced → 3-5 hours in-house. note article = 8-15 hours → 1-2 hours. 30 Threads posts = 15-20 hours → 1-2 hours. Membership site = ¥500K-1M outsourced → 1-3 days. Plus the rate-limit-arbitrage tactic: route to whichever isn't rate-capped, use smux to share the same workspace.
@atelicinvest [Claude Code]
https://x.com/atelicinvest/status/2052622757406339560
Skeptic's take on token-usage hype that's worth keeping. Real-world observations: people burning 50M tokens to build a wedding dashboard, 100M to organize an inbox; Claude 4.7 ignoring a short CLAUDE.md every 5 minutes; 2-3 parallel tasks producing slop because product direction goes incoherent; engineers that audit their own AI output find tons of hallucinated features that the creators are blissfully unaware of. The point: heavy token usage and useful output are not the same thing, and the X-driven narrative skews toward the former.
@Jeanscpa [Claude Code]
https://x.com/Jeanscpa/status/2052600004699955599
A CPA uploaded a new corporate client's past financial statements to Claude Code and asked it to register opening balances in freee (Japanese accounting SaaS). Output: opening balances were registered by counterparty, automatically. The point isn't the workflow — it's that a non-technical accountant can now operate Claude Code as a finance back-office assistant for actual book entries.
@amenya_nelson [Claude Code]
https://x.com/amenya_nelson/status/2052869169469644908
WhatsApp messages now read and replied by Claude Code. The workflow is one of the more obvious utility unlocks of the integration era — most "AI assistant" pitches eventually need to read your messages without leaving them — and the user is operating it. Demo video included.
@0xPhantomDefi [Claude Code]
https://x.com/0xPhantomDefi/status/2052766438037877154
JARVIS built with Claude Code and operated by one clap. Three monitors wake up, Philips Hue switches to focus mode, Spotify starts playing, British voice reads weather/time/daily tasks. Five connected plugins: Wakeup (monitors + voice), Atmosphere (lights + music), Devshop (coding sessions), Project (deadlines/tickets), Mobile (iPhone voice commands). All local through one orchestrator. Monthly cost: ~$640 in API. Estimated savings vs a real assistant: $5K/month.
@leightjessica [Claude Code]
https://x.com/leightjessica/status/2052753644638863535
A development economist used Claude Code to scale up a manual coding exercise: pull all dev econ papers from top general-interest journals + JDE 2020-2025 and code each for whether it's an RCT. Punchline: ~30%, stable. This is the academic-research version of the autoresearch story — the coding work that used to gate empirical reviews is now within reach of one person + a long Claude session.
@sharaff [Claude Code]
https://x.com/sharaff/status/2052768266209341765
Personal expense tracking went fully automatic. Agent watches for BML SMS messages on home network, automatically scans new logs into the expense tracker. No more taking photos of receipts. The "discovery" framing matters: this user just realized that the agent loop closed itself once the SMS-watching skill landed.
@aiedge_ [Claude Code]
https://x.com/aiedge_/status/2052539126646522036
Documents an Obsidian + Claude Code setup as second brain in 10 minutes. Vault = entire business strategy + AI context. Most powerful AI combo claimed. The setup itself is a four-step starter — install both desktop apps, create a vault, write markdown notes/data, paste a connection prompt — but the value is the compounding the vault accrues over months.
🗣 User Voice
User Voice
The day's posts cluster around a few sharper-than-usual signals from heavy users.
Cost ceilings are now the binding constraint. @quxiaoyin: 7B tokens/month across CC + Hermes + Codex. @nicos_ai: free-claude-code proxy via NVIDIA NIM at 40 req/min, Telegram-controlled. @Atenov_D: WindsurfAPI proxy gives Claude Opus 4.7 + GPT-5 free through account pooling. @Axel_bitblaze69: Anthropic-shaped proxy on localhost:8082 routing to NIM/OpenRouter/DeepSeek/Ollama. The Pro plan hits weekly limits before users finish their backlog, and the workaround economy is now bigger than the official integration story.
The HTML-vs-Markdown debate is splitting users. @trq212 makes the structural case (1M context, agent-native rendering). @thehypedotnews counters with the bill: 15x cost per request at $25/M output tokens, and Markdown wins on diffs/portability/speed. @kis: HTML is artifact for humans, Markdown stays for what Claude reads (CLAUDE.md, SKILL.md, README.md). The split that's emerging is structural: HTML for output to humans, Markdown for input to agents.
The Codex vs Claude Code divide is now functional, not tribal. Users running both side-by-side report Codex as the executor, Claude Code as the planner/reviewer. @1osabori: smux to share workspace, rate-limit-arbitrage between the two. @KyleKim84: Codex plugin inside Claude Code, /codex:adversarial-review on plan.md, /codex:review on final code. The teams shipping the most are running parallel sessions with explicit role assignments.
Claude Code's safety harness is regressing. @ClaudeCodeLog flagged 2.1.136 dropped the security-help limiting rule (pentests/CTFs/defensive-only) and the destructive-tactics refusal. Bun HMR disconnect logs are getting injected as top-level prompt sections, replacing parts of "Tone and style." External system-prompt watchers are now functioning as the de-facto regression test for Anthropic's prompt engineering.
The "I built XYZ in Claude Code" story is hitting saturation. @Purbanpp: more Claude Code "experts" on Polish Twitter than people who've actually shipped a project. @atelicinvest: heavy-token-usage influencers tend to be selling skill files, not running real production. The signal-to-noise ratio is degrading, which is why the actual production stories (BridgeMind incident response, the moving-company WhatsApp run, the 47-clients-a-month LP shop) stand out further.
Cost ceilings are now the binding constraint. @quxiaoyin: 7B tokens/month across CC + Hermes + Codex. @nicos_ai: free-claude-code proxy via NVIDIA NIM at 40 req/min, Telegram-controlled. @Atenov_D: WindsurfAPI proxy gives Claude Opus 4.7 + GPT-5 free through account pooling. @Axel_bitblaze69: Anthropic-shaped proxy on localhost:8082 routing to NIM/OpenRouter/DeepSeek/Ollama. The Pro plan hits weekly limits before users finish their backlog, and the workaround economy is now bigger than the official integration story.
The HTML-vs-Markdown debate is splitting users. @trq212 makes the structural case (1M context, agent-native rendering). @thehypedotnews counters with the bill: 15x cost per request at $25/M output tokens, and Markdown wins on diffs/portability/speed. @kis: HTML is artifact for humans, Markdown stays for what Claude reads (CLAUDE.md, SKILL.md, README.md). The split that's emerging is structural: HTML for output to humans, Markdown for input to agents.
The Codex vs Claude Code divide is now functional, not tribal. Users running both side-by-side report Codex as the executor, Claude Code as the planner/reviewer. @1osabori: smux to share workspace, rate-limit-arbitrage between the two. @KyleKim84: Codex plugin inside Claude Code, /codex:adversarial-review on plan.md, /codex:review on final code. The teams shipping the most are running parallel sessions with explicit role assignments.
Claude Code's safety harness is regressing. @ClaudeCodeLog flagged 2.1.136 dropped the security-help limiting rule (pentests/CTFs/defensive-only) and the destructive-tactics refusal. Bun HMR disconnect logs are getting injected as top-level prompt sections, replacing parts of "Tone and style." External system-prompt watchers are now functioning as the de-facto regression test for Anthropic's prompt engineering.
The "I built XYZ in Claude Code" story is hitting saturation. @Purbanpp: more Claude Code "experts" on Polish Twitter than people who've actually shipped a project. @atelicinvest: heavy-token-usage influencers tend to be selling skill files, not running real production. The signal-to-noise ratio is degrading, which is why the actual production stories (BridgeMind incident response, the moving-company WhatsApp run, the 47-clients-a-month LP shop) stand out further.
📡 Eco Products Radar
Eco Products Radar
Tools/products mentioned 3+ times across the day's posts:
Codex / OpenAI Codex (50+) — referenced as the parallel-execution counterpart to Claude Code in dozens of dual-stack workflows.
Cursor (15+) — still the IDE benchmark even when users prefer Claude Code's terminal flow.
Obsidian (20+) — second-brain pattern is now the dominant Claude Code integration target.
OpenClaw (mentioned across both tags, 200+) — the open agent harness, with 2026.5.7 housekeeping release being the day's framing event.
Hermes Agent (30+) — referenced as the compounding-skill alternative when OpenClaw stability fails.
DeepSeek V4 (10+) — $0.14/M tokens vs $5/M for Opus 4.7, now natively speaking the Anthropic API shape.
NVIDIA NIM (5+) — the free-tier model tier behind multiple Claude Code proxies.
Skills / Anthropic Skills (50+) — the SKILL.md format is the unit of reusable agent expertise.
MCP / Model Context Protocol (30+) — the integration layer beneath every connected workflow.
Cowork / Claude Cowork (15+) — desktop counterpart to Claude Code, branding still confused but functional.
Hyperframes / HeyGen / ElevenLabs (10+) — the video-generation triangle behind UGC ad workflows.
Cloudflare (5+) — emerged as the post-AWS-WAF default after BridgeMind incident.
Telegram (15+) — the dominant remote-control channel for cloud agents.
Apify / Google Maps (10+) — the lead-generation feedstock for outbound agent workflows.
freee / Notion / Slack (10+) — the SaaS surfaces that agents now write to directly.
Lovable / Figma Make (5+) — design-side complement to Claude Code in LP/agency workflows.
Karpathy Skills / Superpowers / SuperClaude (10+) — open-source CLAUDE.md frameworks shaping how serious users configure their agents.
Codex / OpenAI Codex (50+) — referenced as the parallel-execution counterpart to Claude Code in dozens of dual-stack workflows.
Cursor (15+) — still the IDE benchmark even when users prefer Claude Code's terminal flow.
Obsidian (20+) — second-brain pattern is now the dominant Claude Code integration target.
OpenClaw (mentioned across both tags, 200+) — the open agent harness, with 2026.5.7 housekeeping release being the day's framing event.
Hermes Agent (30+) — referenced as the compounding-skill alternative when OpenClaw stability fails.
DeepSeek V4 (10+) — $0.14/M tokens vs $5/M for Opus 4.7, now natively speaking the Anthropic API shape.
NVIDIA NIM (5+) — the free-tier model tier behind multiple Claude Code proxies.
Skills / Anthropic Skills (50+) — the SKILL.md format is the unit of reusable agent expertise.
MCP / Model Context Protocol (30+) — the integration layer beneath every connected workflow.
Cowork / Claude Cowork (15+) — desktop counterpart to Claude Code, branding still confused but functional.
Hyperframes / HeyGen / ElevenLabs (10+) — the video-generation triangle behind UGC ad workflows.
Cloudflare (5+) — emerged as the post-AWS-WAF default after BridgeMind incident.
Telegram (15+) — the dominant remote-control channel for cloud agents.
Apify / Google Maps (10+) — the lead-generation feedstock for outbound agent workflows.
freee / Notion / Slack (10+) — the SaaS surfaces that agents now write to directly.
Lovable / Figma Make (5+) — design-side complement to Claude Code in LP/agency workflows.
Karpathy Skills / Superpowers / SuperClaude (10+) — open-source CLAUDE.md frameworks shaping how serious users configure their agents.
Comments