Super User Daily: April 13, 2026
April 11 brought a wave of practitioners refining how they work with Claude Code and OpenClaw, not just using them but engineering better systems around them. The standout theme: persistent memory and context engineering are where the real leverage is. Users who invest in structured knowledge bases and feedback loops are pulling dramatically ahead of those still doing one-shot prompting.
@ianneo_ai [Claude Code]
https://x.com/ianneo_ai/status/2042910729213481143
Implemented Karpathy's LLM Wiki method with Obsidian and Claude Code, creating a three-layer system: raw/ for source materials (read-only), wiki/ for digested notes organized by topic, and trigger words hardcoded into CLAUDE.md. Saying "add to wiki" auto-archives and merges similar items with cascading updates. Saying "lint wiki" runs self-checks for dead links and contradictions. The real workflow change came when writing articles: dump materials into raw/, tell Claude to write based on wiki knowledge (not hallucinated), then auto-generate HTML cards with Playwright screenshots for cover images. The whole article pipeline runs inside Obsidian with full visibility at every step.
@aakashgupta [Claude Code]
https://x.com/aakashgupta/status/2042755527835537814
After iterating on his CLAUDE.md file over 100 times, landed on six sections that actually move the needle: who you are, how you work, your tools, current priorities, skills pointers, and preferences. The compounding trick is what separates good from great: every time Claude Code does something wrong, the fix goes into CLAUDE.md immediately. Over months this file becomes an incredibly precise instruction set. Dave Killeen, CPO at Pendo, told him his system built on this foundation is better than the human executive assistant he used to have. Key mistakes to avoid: making it too long (Claude ignores noise after 100 lines), not updating weekly, and vague preferences like "be helpful."
@HiTw93 [Claude Code]
https://x.com/HiTw93/status/2042933479751716933
Switched all Claude Code conversations to English to improve output quality and personal English skills. Built a custom English Coaching skill into the workflow: after every task execution, the agent outputs corrections on grammar errors, vocabulary choices, and unnatural phrasing. Instead of daily Duolingo sessions, gets real-time language coaching embedded in actual work. The insight is that most AI models have far more English training data, so removing the invisible translation layer improves both the human and the AI output simultaneously.
@iziatask [Claude Code]
https://x.com/iziatask/status/2042910062134296960
Running chatseo.app entirely with Claude Code as the development backbone. Monthly cost breakdown: server at 25 euros, Claude Code at 90 euros, Anthropic API at 2,000 euros, domain and analytics negligible. Total operational cost around 2,100 euros per month for a product serving 7,300 users and generating 11,000 euros MRR. A concrete example of one person building and maintaining a profitable SaaS with AI-first infrastructure where the AI API cost is the dominant expense rather than human labor.
@ishimoto_legal [Claude Code]
https://x.com/ishimoto_legal/status/2042833688921280539
Figured out how to share a single Claude Code instance across multiple PCs by placing the Claude Code folder on Google Drive and pointing each machine at the same path. The one caveat: MEMORY.md accumulates separately on each PC, requiring occasional manual sync. A practical workaround for anyone who works across a desktop at the office and a laptop on the move, eliminating the need to rebuild context on each device.
@milesdeutscher [Claude Code]
https://x.com/milesdeutscher/status/2043079162299371645
Built a three-file persistent memory system that makes Claude remember across sessions. Instructions.md tells the model how to act and includes the crucial line "UPDATE Memory.md with my preferences over time." Memory.md becomes the running brain that auto-updates with corrections and patterns. Context.md holds project-specific context. The real power: saying "stop using em dashes" once gets permanently recorded. The Memory.md file becomes portable across any LLM or Claude chat, eliminating re-explanation overhead on every project.
@leopardracer [Claude Code]
https://x.com/leopardracer/status/2042901993358860335
Used Claude Code to process over 2,000 financial transactions for tax filing. Was originally calculated to owe 13,000 dollars but Claude Code found deductions and categorization opportunities that changed the outcome. A non-coding use case where the value comes from Claude's ability to systematically process large volumes of structured data that would take a human days to review manually.
@kenn [Claude Code]
https://x.com/kenn/status/2042774148867592555
Developed a backlog folder pattern for managing complex Claude Code sessions. When Claude analyzes a system and identifies five important issues, instead of tackling them sequentially in one long context, dump everything as individual markdown files into a backlog/ directory. This enables cross-model review, clean task handoff to new sessions, progressive completion tracking, and follow-up item additions. The habit of "dump everything to markdown" is the single most impactful workflow upgrade for power users.
@noisyb0y1 [Claude Code]
https://x.com/noisyb0y1/status/2042806425596932480
After three months of paying 300 dollars per month for Claude Code, discovered that 70 percent of tokens were being wasted on terminal output noise. Claude was spending 3 dollars per session just reading git status output and test logs. Found two tools that compress command output before feeding it to AI, reducing token consumption from 150,000 to 30,000 per session. Annual cost dropped from 1,200 to 240 dollars with zero change in output quality.
@y_matsuwitter [OpenClaw]
https://x.com/y_matsuwitter/status/2042872424585466206
Delegated daily news collection, X topic monitoring, and email processing to OpenClaw specifically to reduce screen time and internet dependency. The goal is not productivity but reclaiming analog time for pen and paper thinking and physical activity. A counterintuitive use case where the point of deploying more technology is to create space for less technology in daily life.
@oikon48 [Claude Code]
https://x.com/oikon48/status/2042796840865927288
Tested the new /team-onboarding command which analyzes project structure and session usage history to auto-generate an ONBOARDING.md document. The output includes work type breakdown by percentage, top skills and commands ranked by monthly usage, most-used MCP servers with call counts, and a setup checklist for new team members. A practical way to share institutional Claude Code knowledge across a team without writing documentation manually.
🗣 User Voice
User Voice
Token costs and rate limits remain the dominant pain point. Multiple users reported hitting Claude Code limits within 40 minutes of starting work, and several are spending 200 to 300 dollars monthly on subscriptions that still feel restrictive. @0xTengen_ described burning through context windows in under an hour when feeding heavy documentation, despite paying for the Max plan.
Quality regression is a growing concern. @theo noted Claude Code has visibly regressed from earlier versions. AMD AI director @kimmonismus cited analysis of 6,800 sessions showing rising laziness behaviors including shallow reasoning, skipping code review, and incomplete tasks. @youyuxi reported switching to Codex after Claude Code got stuck for 4 minutes on a simple task.
Memory and context persistence remain unsolved at scale. Users like @milesdeutscher and @ishimoto_legal are building manual workarounds with markdown files and Google Drive sync because the native memory system does not reliably carry context across sessions or devices.
Non-engineers face steep onboarding barriers. @dansyu_ican pointed out that all major Claude Code documentation targets engineers, and nobody has systematized workflows for non-technical users. @PandaTalk8 noted the biggest problem for readers is simply getting Claude Code running at all.
The Hermes Agent vs OpenClaw debate intensified. @linyiLYi reported Hermes is more persistent at error recovery and dramatically more token-efficient (context stays at 30-40K tokens vs OpenClaw exceeding 100K), while @lxfater praised Hermes for superior web scraping and memory design but acknowledged being too deeply invested in Claude Code workflows to switch.
Token costs and rate limits remain the dominant pain point. Multiple users reported hitting Claude Code limits within 40 minutes of starting work, and several are spending 200 to 300 dollars monthly on subscriptions that still feel restrictive. @0xTengen_ described burning through context windows in under an hour when feeding heavy documentation, despite paying for the Max plan.
Quality regression is a growing concern. @theo noted Claude Code has visibly regressed from earlier versions. AMD AI director @kimmonismus cited analysis of 6,800 sessions showing rising laziness behaviors including shallow reasoning, skipping code review, and incomplete tasks. @youyuxi reported switching to Codex after Claude Code got stuck for 4 minutes on a simple task.
Memory and context persistence remain unsolved at scale. Users like @milesdeutscher and @ishimoto_legal are building manual workarounds with markdown files and Google Drive sync because the native memory system does not reliably carry context across sessions or devices.
Non-engineers face steep onboarding barriers. @dansyu_ican pointed out that all major Claude Code documentation targets engineers, and nobody has systematized workflows for non-technical users. @PandaTalk8 noted the biggest problem for readers is simply getting Claude Code running at all.
The Hermes Agent vs OpenClaw debate intensified. @linyiLYi reported Hermes is more persistent at error recovery and dramatically more token-efficient (context stays at 30-40K tokens vs OpenClaw exceeding 100K), while @lxfater praised Hermes for superior web scraping and memory design but acknowledged being too deeply invested in Claude Code workflows to switch.
📡 Eco Products Radar
Eco Products Radar
Hermes Agent (Nous Research) — Self-improving AI agent with skill creation loop, competing directly with OpenClaw. Multiple users reporting migration or side-by-side comparison.
Codex (OpenAI) — Frequently mentioned as Claude Code alternative, especially after rate limit frustrations. New 100 dollar Pro plan drawing attention.
Obsidian — Emerging as the default knowledge base for Claude Code power users, especially paired with Karpathy's LLM Wiki method.
RTK — Token compression tool reducing Claude Code costs 60-90 percent by filtering terminal output noise.
Paseo — Remote agent management tool enabling iPhone/iPad control of Claude Code sessions on Mac.
Andrej-karpathy-skills — Single CLAUDE.md file with 4 behavioral principles, hit 12K GitHub stars.
Graphify — Code knowledge graph generator that reduces token consumption 71x per query.
GBrain (garrytan) — Opinionated skill/schema/memory framework for OpenClaw and Hermes Agent.
NotebookLM (Google) — Used as zero-cost research preprocessor to reduce Claude token consumption.
Design.md ecosystem — Growing collection of design system files from 62+ companies for Claude Code UI generation.
Hermes Agent (Nous Research) — Self-improving AI agent with skill creation loop, competing directly with OpenClaw. Multiple users reporting migration or side-by-side comparison.
Codex (OpenAI) — Frequently mentioned as Claude Code alternative, especially after rate limit frustrations. New 100 dollar Pro plan drawing attention.
Obsidian — Emerging as the default knowledge base for Claude Code power users, especially paired with Karpathy's LLM Wiki method.
RTK — Token compression tool reducing Claude Code costs 60-90 percent by filtering terminal output noise.
Paseo — Remote agent management tool enabling iPhone/iPad control of Claude Code sessions on Mac.
Andrej-karpathy-skills — Single CLAUDE.md file with 4 behavioral principles, hit 12K GitHub stars.
Graphify — Code knowledge graph generator that reduces token consumption 71x per query.
GBrain (garrytan) — Opinionated skill/schema/memory framework for OpenClaw and Hermes Agent.
NotebookLM (Google) — Used as zero-cost research preprocessor to reduce Claude token consumption.
Design.md ecosystem — Growing collection of design system files from 62+ companies for Claude Code UI generation.
Comments