# X.com Topic Report: Claude Code Generated at: 2026-04-22T06:44:38.245Z ## Summary - Posts: 8 - Authors: 8 - Tags: N/A - Detail fetched: 4 - Detail not_attempted_detail_limit: 4 ## Topic Digest 共整理 8 篇帖子，来自 8 位作者；暂未提取到稳定高频标签。 ### Top Tags - N/A ### Highlighted Posts - For clarity, we're running a small test on ~2% of new prosumer signups. Existing — Amol Avasare @TheAmolAvasare (2941) https://x.com/decruz/status/2046842523268874375 - How I made Claude Code remember everything across sessions for @VIBEaiRforce Cla — VIBE Giraffe @Vibe_AI_dev (0) https://x.com/Vibe_AI_dev/status/2046842574439133249 - AI with Kyle - Claude Code taken off $20/month Pro Plan? — Kyle Balmer @iamkylebalmer (0) https://x.com/iamkylebalmer/status/2046842561596445078 ### Key Takeaways - For clarity, we're running a small test on ~2% of new prosumer signups. Existing Pro and Max subscribers aren't affected. - How I made Claude Code remember everything across sessions for @VIBEaiRforce Claude Code is smart but stateless. Every session starts from zero. For a project with 389 docs, 50+ custom skills, multi-chain trading agents, and a 689-line CLAUDE.md — that's a disaster. Here's how we fixed it with Obsidian. The problem: - VIBE V2 has 156 items in the project root - 389 markdown files in MD_Docs/ - 80+ docs in docs/ - 21 scattered markdown files at root - Every new Claude session had to rediscover the entire codebase The solution: two Obsidian vaults, one MEMORY.md, zero MCP servers. Vault 1: GigaBrain (GigaBrain) Personal knowledge base. LLM-maintained wiki with sources, entities, concepts, and synthesis pages. The "second brain" for everything Claude architecture research, community patterns, agent design principles. Wikilinks between pages. Auto-loaded CLAUDE.md tells Claude how to ingest, query, and lint the wiki. Vault 2: VIBE-V2-Brain (VIBE-V2-Brain/) Dedicated project knowledge vault. Separate from the code. Contains: -architecture/ — system design, tech stack, data flows - decisions/ — why we chose X over Y - troubleshooting/ — bugs indexed so they never happen twice - integrations/ — CDP SDK, Supabase, Nansen, mem0 patterns - agents/ — tool development, trigger configs, skill inventory - sessions/ — handoff notes between Claude sessions The glue: MEMORY.md in the $VIBE V2 project root. Lightweight. Not a wiki. Just: - Active work (what am I building right now) - Recent decisions (why did I do X) - Blockers (what's stuck) - Session log (what was done each day) Claude Code auto-loads CLAUDE.md every session. We added a section at the top: "MEMORY.md is in the project root read it at the start of every session to continue where you left off. Knowledge vault at /VIBE-V2-Brain/ contains architecture, troubleshooting, integrations. Reference it for deep project knowledge." That's it. No MCP server. No SQLite database. No API calls. Just: 1. Claude starts session → reads MEMORY.md → knows where it left off 2. During work → updates MEMORY.md with decisions and blockers 3. End of session → updates MEMORY.md with what was done 4. Deep questions → point Claude at the VIBE-V2-Brain vault Why separate vaults instead of one: - Code repo = code. Knowledge vault = knowledge. Clean separation. - Obsidian doesn't index node_modules, Python caches, and 156 root items - Each vault has its own git history - Claude Code in VIBE V2 reads MEMORY.md (lightweight). Claude Code in the vault reads deep knowledge (architecture, troubleshooting). No confusion. The real insight: your Obsidian vault IS your CLAUDE.md, but structured, searchable, and compounding. A 689-line CLAUDE.md is a monolith that Claude has to re-read every session. An Obsidian vault with wikilinks lets Claude load only what it needs. Architecture question? - AI with Kyle - Claude Code taken off $20/month Pro Plan? --- ## 1. How I made Claude Code remember everything across sessions for @VIBEaiRforce Cla - URL: https://x.com/Vibe_AI_dev/status/2046842574439133249 - Author: VIBE Giraffe @Vibe_AI_dev - Likes: 0 - Detail capture: fetched > How I made Claude Code remember everything across sessions for @VIBEaiRforce Claude Code is smart but stateless. Every session starts from zero. For a project with 389 docs, 50+ custom skills, How I made Claude Code remember everything across sessions for @VIBEaiRforce Claude Code is smart but stateless. Every session starts from zero. For a project with 389 docs, 50+ custom skills, multi-chain trading agents, and a 689-line CLAUDE.md — that's a disaster. Here's how we fixed it with Obsidian. The problem: - VIBE V2 has 156 items in the project root - 389 markdown files in MD_Docs/ - 80+ docs in docs/ - 21 scattered markdown files at root - Every new Claude session had to rediscover the entire codebase The solution: two Obsidian vaults, one MEMORY.md, zero MCP servers. Vault 1: GigaBrain (GigaBrain) Personal knowledge base. LLM-maintained wiki with sources, entities, concepts, and synthesis pages. The "second brain" for everything Claude architecture research, community patterns, agent design principles. Wikilinks between pages. Auto-loaded CLAUDE.md tells Claude how to ingest, query, and lint the wiki. Vault 2: VIBE-V2-Brain (VIBE-V2-Brain/) Dedicated project knowledge vault. Separate from the code. Contains: -architecture/ — system design, tech stack, data flows - decisions/ — why we chose X over Y - troubleshooting/ — bugs indexed so they never happen twice - integrations/ — CDP SDK, Supabase, Nansen, mem0 patterns - agents/ — tool development, trigger configs, skill inventory - sessions/ — handoff notes between Claude sessions The glue: MEMORY.md in the $VIBE V2 project root. Lightweight. Not a wiki. Just: - Active work (what am I building right now) - Recent decisions (why did I do X) - Blockers (what's stuck) - Session log (what was done each day) Claude Code auto-loads CLAUDE.md every session. We added a section at the top: "MEMORY.md is in the project root read it at the start of every session to continue where you left off. Knowledge vault at /VIBE-V2-Brain/ contains architecture, troubleshooting, integrations. Reference it for deep project knowledge." That's it. No MCP server. No SQLite database. No API calls. Just: 1. Claude starts session → reads MEMORY.md → knows where it left off 2. During work → updates MEMORY.md with decisions and blockers 3. End of session → updates MEMORY.md with what was done 4. Deep questions → point Claude at the VIBE-V2-Brain vault Why separate vaults instead of one: - Code repo = code. Knowledge vault = knowledge. Clean separation. - Obsidian doesn't index node_modules, Python caches, and 156 root items - Each vault has its own git history - Claude Code in VIBE V2 reads MEMORY.md (lightweight). Claude Code in the vault reads deep knowledge (architecture, troubleshooting). No confusion. The real insight: your Obsidian vault IS your CLAUDE.md, but structured, searchable, and compounding. A 689-line CLAUDE.md is a monolith that Claude has to re-read every session. An Obsidian vault with wikilinks lets Claude load only what it needs. Architecture question? Read architecture/system overview.md. Solana swap broke? Read troubleshooting/solana-swap-issues.md. No need to scan 689 lines. The community is converging on this pattern. OpenClaw agents sharing one vault. Claude Code + Obsidian Mind templates. /resume and /wrap-up skills. It's all the same idea: Markdown files as persistent, compounding memory that both humans and LLMs can read. But most people are doing it in one vault. We split it into two because: - Personal knowledge (GigaBrain) and project knowledge (VIBE-V2-Brain) have different lifecycles - A project vault gets messy with code-adjacent noise - GigaBrain connects concepts across projects (agent architecture from Rohit's thread links to our Obsidian pattern links to the LLM Wiki pattern) Three layers of memory: 1. CLAUDE.md — always loaded, rules and gotchas (the law) 2. MEMORY.md — session state, updated every session (the log) 3. Obsidian vaults — deep knowledge, loaded on demand (the brain) No MCP. No servers. No setup beyond creating folders. Build the harness. Then build the infrastructure around it. Then build the memory that makes both survive across sessions. ## 2. AI with Kyle - Claude Code taken off $20/month Pro Plan? - URL: https://x.com/iamkylebalmer/status/2046842561596445078 - Author: Kyle Balmer @iamkylebalmer - Likes: 0 - Detail capture: fetched > AI with Kyle - Claude Code taken off $20/month Pro Plan? AI with Kyle - Claude Code taken off $20/month Pro Plan? ## 3. The 6-month MVP is dead. I'm seeing founders go from idea to payments in 4 hours - URL: https://x.com/alessandro_afl/status/2046842527475531950 - Author: Alessandro Afloarei @alessandro_afl - Likes: 0 - Detail capture: fetched > The 6-month MVP is dead. I'm seeing founders go from idea to payments in 4 hours using Claude Code. Speed is no longer an advantage. It's the baseline. The 6-month MVP is dead. I'm seeing founders go from idea to payments in 4 hours using Claude Code. Speed is no longer an advantage. It's the baseline. ## 4. For clarity, we're running a small test on ~2% of new prosumer signups. Existing - URL: https://x.com/decruz/status/2046842523268874375 - Author: Amol Avasare @TheAmolAvasare - Likes: 2941 - Detail capture: fetched > You really shouldn’t do this. Imagine you’re recommending Claude to someone or your company and you can’t explain why Code suddenly isn’t appearing under the Pro plan. I shouldn’t be the one telling you what the side effects are. For clarity, we're running a small test on ~2% of new prosumer signups. Existing Pro and Max subscribers aren't affected. ## 5. Holy sh*t...Someone built a backend that makes Claude Code 3x cheaper. And it ex - URL: https://x.com/Suryanshti777/status/2046842480038212020 - Author: Suryansh Tiwari @Suryanshti777 - Likes: 0 - Detail capture: not_attempted_detail_limit > Holy sh*t...Someone built a backend that makes Claude Code 3x cheaper. And it exposes a brutal truth: You’re not paying for intelligence. You’re paying for confusion. Same workflow: Bad setup → AI explores like crazy → loops → Optimized setup → AI executes directly ## 6. Claude Code has gotten ridiculously expensive, so I thought, "Maybe I should qui - URL: https://x.com/seedsbiz/status/2046842446810947770 - Author: コスパ アザラシ@重税プラットフォーマー⁨から逃亡中 @seedsbiz - Likes: 0 - Detail capture: not_attempted_detail_limit > Claude Code has gotten ridiculously expensive, so I thought, "Maybe I should quit," but then I found out that CODEX can be installed in CURSOR!!! I can use an AI Agent without changing any environment settings at all!! That's amazing. Everyone, give it a try!! ## 7. 確定申告をClaude Codeで進める手順、3つ。 1. 領収書・通帳・源泉徴収票を1フォルダに集める 形式はバラバラでOK。 PDFでも写真でも。 2. 「 - URL: https://x.com/jikkyo_27/status/2046842403450233321 - Author: 実況中：ベンチャー経営企画の27歳 @jikkyo_27 - Likes: 0 - Detail capture: not_attempted_detail_limit > 確定申告をClaude Codeで進める手順、3つ。 1. 領収書・通帳・源泉徴収票を1フォルダに集める 形式はバラバラでOK。 PDFでも写真でも。 2. 「科目に分けて集計して」と頼む 経費・医療費・寄附金、全部仕分けてくれる。(1/2) ## 8. The difference between Google and Anthropic: Anthropic: Launched Claude Design… - URL: https://x.com/SultanAlFardan/status/2046842393840775480 - Author: سلطان الفردان @SultanAlFardan - Likes: 0 - Detail capture: not_attempted_detail_limit > The difference between Google and Anthropic: Anthropic: Launched Claude Design… and at the same time imposes usage restrictions, even though you can do the same thing via Claude Code Google: Opens DESIGN.md as open source… and lets you work with any tool you want. And the