Read the best tweets about Claude Code, covering terminal workflows, agents, hooks, skills, debugging, and real software projects. Updated every two weeks.
Concrete Claude Code workflows, configuration, production use, and lessons from developers shipping real changes.
Original Xholic analysis
Across 50 tweets from 40 creators, the sampled Claude Code conversation is uniformly favorable (100% SUPPORTIVE stance; 96% POSITIVE sentiment) and skews instructional: TUTORIAL (40%) and LIST (30%) formats dominate. ANNOUNCEMENT-format posts carry the highest median all-time score (109.79), and 78% of posts include media, showing a higher median score (32.64) than text-only posts (18.50). Themes center on architecture, CLAUDE.md project memory, structured workflows, token optimization, and multi-agent orchestration. Note that the sample skews older: median post age is roughly 195 days, with only 4 of 50 posts (8%) published within the last 90 days.
96% of posts
All-time engagement
40% of posts
Published in 90 days
Conversation map
Explanatory breakdowns of Claude Code's building blocks โ skills, hooks, MCP servers, slash commands, plugins, permissions, plan mode, memory โ and how they compose into a system rather than a chatbot.
30%
Disciplined engineering loops: exploreโplanโimplementโcommit, plan mode before edits, TDD red-green-refactor, verification/self-checking (tests, screenshots, browser), requirements grilling (grill-me โ PRD โ issues), and git worktrees.
16%
Writing, structuring, nesting, and maintaining CLAUDE.md (and AGENTS.md) files: templates, hierarchy of global/user/project/local levels, keeping them short, capturing conventions and lessons so context loads automatically each session.
14%
Tips, setups, and workflows shared by Claude Code's creator and Anthropic engineers: parallel sessions, phone-driven workflows, dogfooding new model features (auto mode, recaps, effort levels), prompting workshop, and the self-improving CLA
12%
Bookmark-style roundups of GitHub repos, plugins, MCP servers, and reading resources for extending Claude Code โ orchestration layers, awesome lists, official plugins, best-practice guides.
12%
Diagnosing and fixing recurring agent misbehavior โ silent assumptions, overcomplicated code, unsolicited edits, ignored corrections โ via enforceable config rules (e.g., Karpathy-derived CLAUDE.md principles, claude-doctor transcriptๅๆ)ใ
10%
Reducing token burn and managing context: output-filtering proxies, codebase knowledge graphs, MCP caching/compression, codebase summary files, context assembly internals, and cost/usage analysis tools (ccusage, codeburn).
10%
Scaling beyond one agent: subagents, experimental Agent Teams, parallel Claude sessions, multi-agent swarms and studio-style hierarchies (orca, claude-flow, game-studio repos), and when to use subagents vs. coordinated teams.
8%
Tone and stance
Performance benchmark
Posts with media make up 78% of this collection. Their median all-time score is 32.6, compared with 18.5 for text-only posts.
Format mix
Consensus and debate
Shared view
The dominant framing across the sample is that Claude Code works best as composable engineering infrastructure โ CLAUDE.md memory, hooks, MCP servers, skills, slash commands, and subagents โ rather than a chat interface. This is the top thematic cluster (30% share, 15 tweets) and recurs across tutorial and list posts alike.
Shared view
CLAUDE.md/AGENTS.md content (14% share, median score 26.06) consistently presents the memory file as the mechanism for persisting conventions, lessons, and corrections across sessions, with nesting and brevity as recurring best practices.
Open debate
With 100% of posts coded SUPPORTIVE and 96% POSITIVE, the sampled conversation shows no substantive opposing positions. The closest to internal tension is between posts promoting elaborate multi-repo orchestration setups and posts advising minimal configuration (e.g., building the setup only from repeated real use), though both are framed constructively.
What performs
The 'fixing bad habits' theme has 10% share but disproportionate reach: two of the five engagement outliers sit here. A four-principle CLAUDE.md built from Karpathy's observations about silent assumptions and code bloat (2047550325624594667), the andrej-karpathy-skills repo converting those observations into config rules (2042914348859867218), and a post framing the same four rules as a free fix for Claude Code's bad
Token and context-window content (10% share) posts the strongest theme-level median score (148.55), suggesting cost discipline resonates with practitioners. Examples: a ten-repo round-up of output filters, knowledge graphs, and MCP caches (2045420155434320270), a Tree-sitter codebase graph reported to cut tokens 49x on large monorepos (2041487311536603629), and a script generating five codebase summary files to skip
Posts relaying Boris Cherny's guidance (12% share, median 24.21) are prominent: an account describing his reported setup of parallel sessions, phone-driven mornings, and plan-then-verify loops (2036122417970684148), a CLAUDE.md template the post credits to his team's self-improvement loop (2040356772951544305), and a post relaying his dogfooded Opus 4.7 features โ auto mode, recaps, effort levels โ and his reported /
Eight percent of posts cover scaling beyond one agent: experimental Agent Teams distinguishing coordinating teammates from fire-and-forget subagents (2035047653583565121), a hackathon-winning agent/skill/command stack with red-team pipelines (2071602594318008774), and effort-mode-driven automatic subagent orchestration (2080046477322105256). Its modest theme median (19.02) and the explicit 'still experimental' caveat
ANNOUNCEMENT-format posts carry the highest format median (109.79) โ official mod support shipping inside plugins (2105721434807083061), the open-sourced 27-agent/64-skill setup (2038660653410320556), and the andrej-karpathy-skills repo (2042914348859867218) โ versus 18.50 for OPINION. With 78% of posts using media (median 32.64 vs 18.50 text-only), visual tutorials and lists are the house style. Curated repo lists (
Statistical standouts
Creator landscape
The five most represented creators account for 20% of the selected posts.
1. Vaishnavi
@_vmlops
2 posts
2. Dharmik Harinkhede
@AiwithDharmik
2 posts
3. CyrilXBT
@cyrilXBT
2 posts
4. Harshil Tomar
@Hartdrawss
2 posts
5. Jian Wang
@jianw851
2 posts
6. 0xMarioNawfal
@RoundtableSpace
2 posts
The top thematic cluster (30% share, 15 tweets) explains Claude Code's building blocks โ skills, hooks, MCP servers, slash commands, plugins, permissions, plan mode, and memory โ as composable engineering infrastructure rather than a chat interface. Examples include the official mod/plugin announcement (2105721434807083061), a guide claiming 16 of 93 commands deliver 80% of the value (2036843438193496526), and a six-
Seven posts (14% share, median score 26.06) treat CLAUDE.md/AGENTS.md as the system's keystone: nested per-directory files that load context automatically up the folder tree (2044778874580656486), a template the post credits to Boris Cherny's team that captures mistakes as permanent rules (2040356772951544305), and AGENTS.md discipline โ plan first, smallest change, verify, log corrections โ so the file improves over
Structured engineering workflows (16% share, 8 posts) recur as a counterweight to ad-hoc prompting: exploreโplanโimplementโcommit with TDD and git worktrees (2038917245506871764), cross-model review by having Codex audit Claude's plan and code (2040497640174137443), and a grill-meโPRDโissuesโTDD pipeline that forces requirements resolution before code (2072281633818832995).
Since the previous snapshot
Themes, sentiment, stance, and post format are classified per tweet. All counts, shares, medians, creator concentration, freshness, and performance comparisons are then calculated directly from the published snapshot.
Xholic's all-time score compares engagement while accounting for reach, post age, and creator consistency. It is used for relative comparisons within this collection.
This report analyzes the exact 50-post snapshot shown below. AI identifies editorial categories and drafts explanations; all statistics are calculated from the snapshot, and every narrative claim is checked against cited posts before publication.
Best Claude Code tweets
Ranked 01โ50
@dr_cintas ยท
This is the most complete Claude Code setup that exists right now. 27 agents. 64 skills. 33 commands. All open source. The Anthropic hackathon winner open-sourced his entire system, refined over 10 months of building real products. What's inside: โ 27 agents (plan, review, fix builds, security audits) โ 64 skills (TDD, token optimization, memory persistence) โ 33 commands (/plan, /tdd, /security-scan, /refactor-clean) โ AgentShield: 1,282 security tests, 98% coverage 60% documented cost reduction. Works on Claude Code, Cursor, OpenCode, Codex CLI. 100% open source.
@om_patel5 ยท
SOMEONE BUILT A SINGLE CLAUDE.MD FILE THAT FIXES EVERY BAD HABIT CLAUDE CODE HAS AND IT HIT 78.5K STARS it's based on andrej karpathy's public observations about how LLMs write code the problem he pointed out is that claude makes silent assumptions, overcomplicates everything, writes 1000 lines when 100 would do, and sometimes deletes code it doesn't fully understand as a side effect so forrestchang turned karpathy's critique into 4 behavioral principles and dropped them in one claude.md file: 1\ surface your assumptions don't pick an interpretation silently. if there are multiple ways to read the task, say so. if uncertain, ask. push back when something doesn't make sense instead of just running with a bad plan 2\ minimum viable code no speculative features, no abstractions for single use code, no "flexibility" you weren't asked for. if you wrote 200 lines and 50 would work, REWRITE IT. ask yourself if a senior engineer would call this overcomplicated 3\ surgical changes only don't touch code you don't fully understand, don't refactor unrelated stuff as a side effect, don't delete comments because they look unnecessary. only change what the task actually requires 4\ goal driven execution give claude success criteria instead of step by step instructions. karpathy's exact quote: "LLMs are exceptionally good at looping until they meet specific goals. don't tell it what to do, give it success criteria and watch it go" one file has 78.5k STARS AND 7.4k FORKS on a single github repo. install is one curl command that drops it straight into your ~/.claude folder
@sharbel ยท
๐จ Andrej Karpathy documented the exact ways LLMs fail at coding. Someone turned those observations into a single Claude config file. It's called andrej-karpathy-skills. +3,741 stars this week. Why it's great: Claude Code makes the same mistakes on every project. It over-explains. It adds code you didn't ask for. It ignores constraints you set 3 prompts ago. Most people just accept this as the baseline. Karpathy didn't. He catalogued the failure patterns. This repo converts every one of them into a CLAUDE.md instruction that fixes the behavior at the source. How to use it: Drop the CLAUDE.md file into the root of any project. Claude reads it automatically on every session. No prompt engineering on every request. No babysitting. The behavior changes once and stays changed. One file. Every project.
@ClaudeDevs ยท
You can now mod Claude Code: - Change how it behaves - Customize the UI - Swap in your own features Write one with a few lines of TypeScript, or have Claude build it for you. Mods ship inside plugins, so you install them with /plugin in the CLI or desktop app. A few examples:
Watch video
@startupideaspod ยท
Boris Cherny, the creator of Claude Code, shared his entire setup. He runs 5-10 Claudes in parallel. Half his coding happens from his phone. Here's his 3-part formula for better results: Use the smartest model available โ Counterintuitive: it's actually cheaper โ Smarter model = fewer tokens = lower total cost โ "Once the plan is good, the code is good" Invest in your Claude MD โ Plain text file. No special format. โ Whole team contributes multiple times a week โ Every mistake Claude makes gets added so it never happens again Give Claude a way to verify its own output โ Let it run the code. Let it see the browser. โ "Imagine you're a painter wearing a blindfold" โ Same thing for an AI that can never check its work His morning routine: wake up, kick off 3 sessions from his phone, check in later. His workflow: start in plan mode โ lock the plan โ auto-accept edits โ done. No fancy setup. No complex tooling. Just multiple Claudes, a good plan, and a shared knowledge base.
Watch video
@DeRonin_ ยท
10 GitHub repos to spend 60-90% less tokens in Claude Code: 1. RTK (Rust Token Killer) CLI proxy that filters terminal output before it hits your context - 60-90% reduction on common dev commands - one binary, zero dependencies - works with Claude Code, Cursor, Copilot Repo: https://t.co/WayvpBtyBH 2. Context Mode Sandboxes raw tool output into SQLite instead of dumping it into context - 98% context reduction on Playwright, GitHub, logs - only clean summaries enter your conversation - works as Claude Code plugin Repo: https://t.co/YNbFIGQz7X 3. code-review-graph Local knowledge graph that maps your codebase with Tree-sitter - Claude reads only what matters, not the entire repo - 49x token reduction on large monorepos - 6.8x on average reviews Repo: https://t.co/9gIzmAWN12 4. Token Savior MCP server that navigates code by symbols, not full files - 97% reduction on code navigation - persistent memory across sessions - 69 tools, zero external deps Repo: https://t.co/OtvhrMgGWh 5. Caveman Claude makes Claude talk like a caveman to cut output tokens - 65-75% output reduction - one-line install - keeps full technical accuracy Repo: https://t.co/onBeghTyfH 6. claude-token-efficient one CLAUDE.md file that keeps responses terse - drop-in, no code changes - reduces output verbosity on heavy workflows - best for output-heavy sessions Repo: https://t.co/j6MKo9klQe 7. token-optimizer-mcp MCP server with caching, compression, and smart tool intelligence - 95%+ token reduction through intelligent caching - compresses repeated tool outputs Repo: https://t.co/0jIVQ4ANls 8. claude-token-optimizer reusable setup prompts for optimizing any project - 90% token savings in 5 minutes - reduces doc token usage from 11K to 1.3K Repo: https://t.co/puil9WwFGB 9. token-optimizer finds ghost tokens that silently eat your context - survives compaction without losing quality - fixes context quality decay Repo: https://t.co/92G8e4yeGq 10. claude-context (by Zilliz) code search MCP that makes your entire codebase the context - ~40% reduction with equivalent retrieval quality - hybrid BM25 + dense vector search Repo: https://t.co/yjfiQOSy15 [ how to stack them ]: you don't need all 10. pick 2-3 based on your workflow: > heavy terminal output? RTK > big codebase? code-review-graph + Token Savior > lots of MCP servers? Context Mode > quick fix? Caveman + claude-token-efficient most people are burning tokens without knowing it run /context in a fresh session and see how much is gone before you even type a word your pocket will thank me later :<)
@jianw851 ยท
Claude Code has 93 commands. You only need these 16 to get 80% of the value ๐ /init โ generates CLAUDE.md (your repo memory) /plan โ think first, code later /context โ shows whatโs eating your tokens /compact โ free up context mid-session /clear โ reset between tasks /model โ switch Opus / Sonnet / Haiku /btw โ ask side questions without breaking flow /rewind โ undo when things go off track /agents โ run multiple sub-agents in parallel /chrome โ control browser (click, fill, navigate) /loop โ run prompts on repeat (monitoring) /simplify โ 3 agents review your code together /permissions โ pre-allow safe actions --dangerously-skip-permissions โ full auto mode (no interruptions) Shift + Tab โ switch modes instantly "ultra think" โ max reasoning depth Most people use Claude Code like a chatbot. Thatโs why they struggle. Power users run it like a system. Save this. Youโll need it. โป๏ธ #AI #Claude #AIAgents #LLM #GenAI #DevTools
@hasantoxr ยท
๐จBREAKING: Someone built a local knowledge graph for Claude Code that cuts token usage by 49x on daily coding tasks. It's called code-review-graph and it builds a persistent structural map of your entire codebase using Tree-sitter so Claude reads only the files that actually matter instead of burning tokens scanning everything. โ 8.2x average token reduction across 6 real repositories โ Blast-radius analysis traces every caller, dependent, and test affected by any change โ Incremental updates re-parse only changed files in under 2 seconds โ Works with Claude Code, Cursor, Windsurf, Zed, and more out of the box 100% Opensource. Link in comments.
@sickdotdev ยท
Everyone is talking about Claude Code. Very few people actually understand it. The ones who do aren't smarter. They just have a mental model that the hype crowd doesn't. Here's the model: Claude Code has 6 core components. Once you internalize them, the whole thing clicks. You stop fumbling and start building with intention. โโโโโโโโโโโโโโโโโโโโ ๐. CLAUDE. md โ Project Memory This is where Claude learns your project before touching a single file. Conventions, architecture, rules. Without this, you're prompting blind every session. ๐. Hooks โ Lifecycle Control Deterministic shell commands at exact lifecycle points. Before a tool runs. After a file is written. When a session ends. This is how serious teams enforce standards instead of hoping Claude remembers. ๐. MCP Servers โ External Integrations This is what separates a coding assistant from an actual agent. Claude connects to GitHub, Slack, databases, internal APIs โ and acts on them. Not just code generation. Real workflow execution. ๐. Subagents โ Parallel Orchestration One Claude orchestrates. Others execute in parallel โ research, implementation, testing โ all running at the same time. This is where the productivity ceiling breaks open. ๐. Skills โ Reusable Knowledge SKILL. md files encode domain expertise Claude reads before every relevant task. You teach it once. It applies that knowledge forever. Your best decisions become the default. ๐. Computer Use โ Full Execution Claude doesn't just write code. It navigates browsers, fills forms, controls desktop UI. The loop closes. Idea to shipped โ without you touching the keyboard. โโโโโโโโโโโโโโโโโโโโ Most people will keep using Claude Code as a fancy autocomplete. The ones who go deep into each of these 6 components will build things that look impossible to everyone still lost in the hype. The architecture is the advantage. Which of these 6 do you think is the most underrated โ and why?
@TaylorPearsonMe ยท
I think the top tip I've picked up since I started using Claude Code: write a CLAUDE [dot] md at every level of your file structure and be intelligent about how you nest folders. Claude Code automatically loads the Claudemd file from every directory level above whatever file you're working on. So if you organize it intelligently, it always has the right context at the right time. Say you're running a business and doing some marketing work. You might have: - A general CLAUDEmd on your computer โ how you like Claude to work, voice, tools you have wired up - A vault CLAUDEmd โ how your files are organized - A folder for your business with a CLAUDEmd โ what the business does, who the clients are, how you price, etc. - A folder for marketing inside the business with a CLAUDEmd โ channels you run, what's working, brand voice, etc. - The specific marketing project with its own CLAUDEmd When you open anything in the marketing project folder, Claude Code walks up the tree and pulls all five into context automatically. Global โ vault โ business โ marketing โ project. That means the very first message in a chat about that marketing project is going to be like talking to someone that knows you, your business, your marketing history, etc. You need to keep the CLAUDEmd at each level current for this to work I have a "wrap" skill that I run at the end of every chat session. It updates the relevant CLAUDEmd(s) and makes a more detailed entry into a workbench file for whatever I was working on. So if I make a decision about a project or learn something about a client, that context lives in the right file by the time I close the session. Next time I come back, Claude already has it. Starter repo for this kind of setup: https://t.co/m78XGeG7ii I built a little mini-extension for the workbench structure I use: https://t.co/dKQPab0EJv
@jianw851 ยท
Most developers are using Claude Code wrong. They treat it like a coding assistant: Prompt โ Output โ Repeat. That worksโฆ until it doesnโt. Because real AI development isnโt about prompts. Itโs about systems. Hereโs the Claude Code setup that changes everything ๐ โข CLAUDE.md โ the brain (context, rules, instructions) โข /skills โ reusable workflows (review, refactor, release) โข /hooks โ guardrails + automation โข /docs โ architecture decisions (the โwhyโ) โข /src โ actual product logic This isnโt just folder structure. Itโs how you: โข stop repeating instructions โข get consistent outputs โข scale across features + teams โข turn AI into a predictable system Most devs keep rewriting prompts. Better ones design systems where AI doesnโt need reminders. Thatโs when Claude stops guessingโฆ โฆand starts acting like an engineering partner. Save this. โป๏ธ #AI #Claude #AIAgents #LLM #GenAI #DevTools
@VaibhavSisinty ยท
Someone on Reddit built a single script that saves 50,000 tokens per Claude Code conversation. Every time you start a conversation, Claude Code spends 10-20 tool calls just exploring your codebase. On large projects, 30-50K tokens gone before any real work begins. The fix is called ai-codex. Run it once and it generates 5 compact markdown files: โ routes.md : every API route with methods and auth tags โ pages.md : full page tree with client/server flags โ lib.md : all library exports with function signatures โ schema.md : database schema compressed to key fields only โ components.md : component index with props Add one line to your CLAUDE.md. Claude reads these first. Exploration phase gone.
@Suryanshti777 ยท
๐จBreaking: The guy who created Claude Code (@bcherny) just revealed how his team actually trains their AI. One file: CLAUDE.md You place it at the root of your project. Inside it: past mistakes conventions rules Claude reads it every session. The result? The agent improves over time without you touching the code. Every bug that gets fixed becomes a permanent rule. Boris Cherny uses this internally at Anthropic every day. Hereโs the template he shared โ ready to copy, paste, and adapt. CLAUDE.md Template 1. Plan Mode Default Enter plan mode for any non-trivial task (3+ steps or architectural decisions) If something goes wrong, STOP and re-plan immediately โ donโt keep pushing Use plan mode for verification steps, not just building Write detailed specs upfront to reduce ambiguity 2. Subagent Strategy Use subagents frequently to keep the main context window clean Offload research, exploration, and parallel analysis to subagents For complex problems, throw more compute via subagents Assign one task per subagent for focused execution 3. Self-Improvement Loop After any correction from the user, update tasks/lessons.md with the pattern Write rules for yourself to prevent repeating the same mistake Ruthlessly iterate on these lessons until the mistake rate drops Review lessons at the start of each session 4. Verification Before Done Never mark a task complete without proving it works Diff behavior between main and your changes when relevant Ask yourself: โWould a staff engineer approve this?โ Run tests, check logs, and demonstrate correctness 5. Demand Elegance (Balanced) For non-trivial changes, ask: โIs there a more elegant solution?โ If a fix feels hacky, ask: โKnowing everything I know now, implement the elegant solution.โ Skip this for simple fixes โ donโt over-engineer Challenge your own work before presenting it 6. Autonomous Bug Fixing When given a bug report: just fix it Use logs, errors, and failing tests to diagnose Require zero context switching from the user Fix failing CI tests automatically Task Management 1. Plan First โ Write the plan in tasks/todo.md with checkable items 2. Verify Plan โ Confirm the plan before implementation 3. Track Progress โ Mark items complete as you go 4. Explain Changes โ Provide a high-level summary at each step 5. Document Results โ Add a review section to tasks/todo.md 6. Capture Lessons โ Update tasks/lessons.md after corrections Core Principles Simplicity First Make every change as simple as possible and minimize code impact. No Laziness Find root causes. Avoid temporary fixes. Maintain senior-level engineering standards.
@RoundtableSpace ยท
one AGENTS.md file can make Claude Code and Codex work a lot more like a disciplined engineering team. Codex reads it natively, and recent Claude Code versions can use it when thereโs no CLAUDE.md. the structure is simple: โ plan before touching code โ make the smallest change that works โ split work across subagents โ own the bug โ verify before saying done โ write down every correction the best part: the file gets better over time as you keep adding lessons from real mistakes. Claude Code users can also install the official setup plugin: `/plugin install claude-code-setup@claude-plugins-official` then have it scan your codebase and recommend useful hooks, skills and subagents. copy the AGENTS.md below.
@cyrilXBT ยท
This is INSANEE... Claude Code is one engineer. These 7 repos turn it into a whole team: THE MANAGER Orca โ runs multiple agents in parallel https://t.co/jysYX2z7T7 THE RESEARCHER XERJ โ local search index, 40x fewer tokens than grep https://t.co/pNlzXeZn27 THE MEMORY Context Mode โ shrinks tool output by up to 98% https://t.co/QZh1fkM8RD THE SENIOR DEV Matt Pocock's Skills โ battle tested coding skills https://t.co/yJ89h6jYGI THE MIGRATION EXPERT Code Modernization โ official Anthropic plugin https://t.co/YkHrsM0KPh THE SPECIALIST Claude Code MCP โ an agent inside your agent https://t.co/76hy9HdnID THE ONBOARDING DOC Claude Code Ultimate Guide https://t.co/hS4siiYlbj Bookmark this.
@JJEnglert ยท
Our engineers at @tenex_labs are slamming this new Claude Code feature. Here's how it works: @AnthropicAI just shipped Agent Teams โ and it's a big deal even if you're not writing code yourself. Here's the simple version: Instead of one AI working on your problem alone, you can now spin up a whole team of AI agents that work together. One acts as the lead. The others are teammates. They each focus on a different piece of the work, talk to each other directly, and coordinate through a shared task list. Think of it like hiring a project manager who breaks the work into pieces, assigns it to specialists, and makes sure nothing falls through the cracks. Except all of those people are Claude, and they spin up in seconds. Why this matters even if you're not a developer If you've ever used Claude Code to research a problem, write a report, or automate a workflow โ you've been working with one brain at a time. That brain has a limit on how much it can hold in its head before things start slipping. Agent Teams removes that bottleneck. Each teammate gets its own full memory. One can be deep in your financials while another is reviewing your competitor landscape while a third is drafting recommendations. They don't confuse each other's work because they literally can't see each other's context. And the best part โ they talk to each other. The lead doesn't have to relay everything. Teammates share findings, challenge each other's conclusions, and build on each other's work directly. "Wait, didn't Claude Code already have subagents?" Yes. And this is where most people get confused. Here's the difference: Subagents are like sending an intern to go research something and come back with an answer. They do the work, hand you a summary, and they're done. They never talk to each other. You manage everything. Simple, cheap, good for focused tasks where you just need a result. Agent Teams are like assembling a working group. The teammates coordinate with each other, not just with you. They claim tasks from a shared list, message each other when they find something relevant, and the lead synthesizes everything at the end. More expensive, but the output is fundamentally different. When to use which: - Need a quick answer or a focused task done? Subagent. Fast, cheap, gets the job done. - Need multiple people looking at different angles of the same problem? Agent Team. The coordination is the point. - One person can do it without talking to anyone else? Subagent. The work benefits from debate, cross-checking, or parallel exploration? Agent Team. What our team is using it for right now: 1. Parallel code reviews โ 3 teammates reviewing the same PR simultaneously. One on security, one on performance, one on test coverage. A single reviewer gravitates toward one issue type. Three specialists catch everything. 2. Competing hypotheses โ 5 agents investigating the same bug, each with a different theory, actively trying to disprove each other. The theory that survives is almost always the root cause. 3. Cross-layer features โ Frontend, backend, and tests each owned by a different teammate. No one steps on anyone else's work. Quick start if you want to try it: Requires Claude Code v2.1.32 or later. Add one environment variable to your settings.json: "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1" Tell Claude what kind of team you want in plain English. "Create a team with 3 teammates to review this project from different angles." Claude proposes the team, you confirm, and it handles the rest โ spawning teammates, assigning tasks, coordinating work. Start with 3 teammates. Keep tasks independent. Don't let two teammates touch the same files. Still experimental. But this is the first multi-agent architecture I've seen actually hold up on real dev work. Link to docs in comments below.
@cyrilXBT ยท
Claude Code can build full apps. These 8 repos make it build them better. PLAN IT gstack โ Garry Tan's setup that gives Claude Code roles like founder, eng manager and QA https://t.co/G2C5IJQq6O STAFF IT claude-flow โ runs swarms of agents working together https://t.co/eQQsm7YfI4 oh-my-claudecode โ multi agent orchestration on top of Claude Code https://t.co/MM6wevLBRt orca โ parallel agents https://t.co/jysYX2z7T7 REMEMBER IT claude-mem โ memory that survives between sessions https://t.co/1XI7Odux4C context-mode โ keeps tool output from flooding the context https://t.co/QZh1fkM8RD DO IT RIGHT agent-skills โ senior engineer habits: specs, TDD, reviews https://t.co/bONo0s9NJu everything-claude-code โ agents, skills, hooks and configs in one place https://t.co/2Er9PREAih Bookmark this.
@TaylorPearsonMe ยท
A friend asked me what to read to get started with Claude Code and here was my attempt at the most practical resources I've found: (This list is AI-generated, as are the descriptions, but I've read all of these and curated the selection. More on how this was made at the bottom b/c it's an interesting example of an AI workflow that's much better than how this would have worked in the past.) Getting started: 1. Best Practices for Claude Code (Anthropic docs) โ the single best starting point. Covers CLAUDE dot md, prompting, hooks, skills. https://t.co/AUWTK0tF3V 2. CodeLife: How I Use AI (@DaveNadig) โ clear explanation of the Claude Code loop (observe-orient-decide-act) and how it differs from Cursor-style autocomplete. https://t.co/iPA9Vk47s6 Getting better: 3. Claude Code Customization: When to Use Rules, Skills, Subagents, and MCPs (@maboroshi_mario) โ practical decision tree for each customization layer. https://t.co/GKKvbSAIBk 4. Tips from the creator (@ABorisC) โ Boris Cherny's own tips, including learning mode in /config and having Claude generate visual presentations. https://t.co/DHgExGmLUh 5. How We Use Skills (@trq212, Anthropic) โ how to structure skills so Claude spends turns on composition, not boilerplate. "The gotchas section is the highest-signal content in any skill." https://t.co/eQxprAVepN Mental models: 6. Management as AI Superpower (@emollick) โ "the ability to specify what good looks like" is the bottleneck. Delegation skills transfer directly. https://t.co/kYPVUadMqD 7. Compound Engineering (Every) โ why building up CLAUDE.md, skills, and commands pays compound returns. Each session makes the next one more productive. https://t.co/jjsTeldRap 8. The Magic of Claude Code (@fromedome) โ why filesystem access + Unix philosophy makes Claude Code fundamentally different from chatbots. The mental model for understanding what you're actually working with. https://t.co/l3mpm6QmHE --- How this list was made: I use @Readwise Reader as my read-it-later app (it's great!). Everything I save (articles, tweets, podcasts) gets synced into my Obsidian vault along with my highlights and notes. I also write a monthly links newsletter (The Interesting Times), where I curate and annotate the best things I've read. So when my friend asked what to read, I already had the primitives sitting there: two months of saved articles with highlights, plus newsletter drafts where I'd already written descriptions of the best ones. I asked Claude Code to search my Readwise library and newsletter drafts for the most practical Claude Code resources, pull one- or two-sentence descriptions (mostly from how I'd already described them in the newsletter), and compile a ranked list. The whole thing took maybe five minutes. Historically this would have been 30-45 minutes of me scrolling through my reading history trying to remember what was good, re-reading intros to jog my memory, and writing descriptions from scratch. This is a small example of natural language composability. Because Claude Code can read your files, pull from external sources like Readwise, and reference things you've already written, you're constantly building on top of previous work. The newsletter descriptions I wrote last month became the raw material for this list today which only took five minutes to put together. Then, of course, I just told Claude to turn it into a tweet and write a description of how it was put together, which is what you are reading now. (I do manually read and edit everything though before sending thouhg.)
@_avichawla ยท
12 must-use features in Claude Code: - CLAUDE .md is your project's memory. It stores your stack details, conventions, and rules so Claude loads them at every session start. - Permissions let you whitelist or block tools like Bash per session. If you're working on anything production-facing, this is non-negotiable. - Plan Mode makes Claude draft a step-by-step plan before touching any code. You get to approve or reject before anything runs. - Rules let you set project-wide behavioral guardrails with specific do's and don'ts beyond what CLAUDE(.)md covers. - Skills are reusable instruction files you store in .claude/skills/. Write them once and Claude follows them automatically every time. - Hooks fire shell scripts on events like PreToolUse and PostToolUse, which makes them perfect for auto-linting or triggering tests. - MCP connects Claude to databases, APIs, and services. This is how you give it real-world access beyond your codebase. - Plugins let you add Docker, pytest, and VS Code extensions without writing any integration code. - Slash Commands store workflow shortcuts in .claude/commands/ so you can trigger complex flows with a single keystroke. - Subagents spawn parallel Claude instances that divide and conquer multi-step workflows simultaneously. - Voice Mode lets you talk to Claude hands-free, which is great for quick queries while your hands are on the keyboard. - Rewind lets you step back to any checkpoint in your session and restart cleanly from that point. ๐ Over to you: Which features do you use the most in CC? ByteByteGo put out a great post on Claude Code features every engineer should know. I took a different angle here and focused on the ones you should actually be using.
Watch video
@heygurisingh ยท
Holy shit... someone built a TUI that shows where your Claude Code tokens actually go. Turns out 56% of my $200/day spend wasn't coding. It was Claude just... talking. It's called codeburn. Reads the session transcripts Claude Code already stores at ~/.claude/projects/ and classifies every turn into 13 categories based on tool usage patterns. No LLM calls for classification. Fully deterministic. No config. The breakdown that broke me: >> Conversation (no tools used): 56% >> Actual coding (edits + writes): 21% >> Everything else: 23% I was paying $200/day to watch Claude type paragraphs at me. ccusage shows cost per model per day. codeburn shows cost per task type, project, model, tool, and MCP server. Daily activity chart with gradient bars. Arrow keys to switch between today / week / month. SwiftBar menu bar widget if you're on Mac. One command. Works with any Claude Code install: npx codeburn
@AiwithDharmik ยท
Most developers treat Claude Code like a smarter autocomplete. That's the wrong mental model. It's actually a 4-layer engineering system: 1๏ธโฃ CLAUDE.md โ persistent project memory Architecture, rules, team conventions 2๏ธโฃ Skills โ auto-invoked knowledge packs Testing patterns, code review, deploy workflows 3๏ธโฃ Hooks โ deterministic guardrails Security checks, formatting, automation 4๏ธโฃ Agents โ specialized sub-agents Break complex tasks into parallel workflows Once these are configured properly: Claude stops behaving like a chatbot. It starts behaving like a senior engineer on your team. Most people never reach this level because they skip the setup. The gap between average AI output and production-level results isn't the model. It's the infrastructure around it. Hereโs the full breakdown on exactly how to build this ๐
@AiwithDharmik ยท
You donโt need another Claude Code setup. Just this one. Everyone kept asking: โ Where do Skills go? โ How do Hooks work? โ How do you connect MCP servers? So hereโs the definitive Claude Code structure๐ Core layout โข CLAUDE.md โ your projectโs memory layer โข .claude/ โ extensions hub โข commands/ โ reusable slash commands โข skills/ โ auto-triggered workflows (SKILL.md) โข agents/ โ sub-agents (.yml) โข plugins/ โ packaged setups โข .mcp.json โ external tool connections What most people miss This isnโt just folders. Itโs a system: โข 6 extension types (how everything plugs in) โข Hook events โ when + how automation runs โข Skill structure โ scripts, refs, assets โข MCP servers โ real tool integrations โข Context limits โ when Claude starts dropping info โข Setup flow โ from zero โ production Most people use Claude Code like a chatbot. Thatโs why it breaks at scale. This setup turns it into a real engineering system. Share it. Bookmark it. Use it. This is the only structure youโll need. #AI #Claude #AIAgents #LLM #GenAI #DevTools
@techNmak ยท
10 GitHub repos that actually earn their place in a Claude Code workflow : 1./ superpowers โญ142K TDD-enforced. Deletes code written before tests exist. Forces structured thinking before any line runs. https://t.co/gUpoHVN3E6 2./ everything-claude-code โญ147K 30 agents, 136 skills, 60 slash commands. Don't install all of it, pick 4. Still changes everything. https://t.co/UnISd5KCqX 3./ repomix โญ23K One command packs your entire codebase into one AI-readable file. Claude sees the full picture instantly. https://t.co/1XKUO5K8xO 4./ awesome-claude-code โญ37K The community's curated index. Only PR allowed: from Claude itself. That's the quality bar. https://t.co/ptpZVOdLgx 5./ claude-code-action โญ7K โ Official Runs Claude Code directly in your CI/CD pipeline. PR reviews, automated, not demo-ware. https://t.co/HIMqZ5doIv 6./ claude-code-best-practice โญ33K Read it, don't install it. Apply the patterns. 30 mins โ weeks saved. https://t.co/93ixBVxVVg 7./ servers (GitHub MCP) โญ83K Full GitHub access via natural language. Read repos, create issues, manage PRs. https://t.co/ddH1nK4tEa 8./ context7 โญ52K Injects live, up-to-date library docs into your prompts. Kills hallucinated APIs dead. https://t.co/o5ujwq8hhc 9./ claude-session-restore Restores full context from previous sessions via git history. Handles files up to 2GB. https://t.co/KsWvGcjHp0 10./ awesome-claude-skills โญ52K 50+ community skills by category. Browse before you build your own. https://t.co/mgebMQT29H
@PrajwalTomar_ ยท
A solo dev won $15,000 at Anthropic's hackathon in 8 hours, then gave the entire setup away for free. It's the clearest masterclass I've seen in running Claude Code like real infrastructure. I've been telling my clients that this is how the people pulling ahead actually use it. This is the proof. Here's what's in the setup: 1. 38 specialized agents. Planner, security reviewer, debugger, code reviewer. Each with one job instead of one agent doing everything badly. 2. 156 skills loaded on demand. /plan, /tdd, /security-scan, /quality-gate. It pulls the right skill for the task instead of guessing. 3. 72 custom slash commands. Repeatable workflows turned into one keystroke. 4. AgentShield: 1,282 security tests across CLAUDE .md, MCP configs, hooks, and skills. Most people ship agents with zero security checks. 5. 3 Opus 4.6 agents running red-team pipelines. Attacker, Defender, Auditor. They stress-test each other before anything ships. 6. A continuous learning layer that builds confidence across sessions, so it gets better the longer it runs. 7. Coverage across 12 language ecosystems. The lesson isn't the repo. It's the mindset. The builders pulling ahead treat Claude Code like infrastructure they engineer, not something they prompt one line at a time. If you're barely scratching the surface with Claude Code, this is your wake-up call. Bookmark this. โ
@_vmlops ยท
THE CREATOR OF CLAUDE CODE JUST SHARED 6 TIPS AFTER DOGFOODING OPUS 4.7 FOR WEEKS Boris Cherny built claude code. and after weeks of using opus 4.7, he dropped 6 tips that actually change how you work with it here's what's new: โ auto mode โ no more babysitting claude on long tasks. it decides what's safe to run and what needs your approval...run multiple claudes in parallel now โ /fewer-permission-prompts โ scans your session and builds a safe allowlist...stop clicking "approve" for the same commands every time โ recaps โ claude summarizes what it did and what's next. come back after hours and instantly know where you left off โ focus mode (/focus) โ hides all the intermediate noise. just see the final result...boris says he trusts the model enough now to not watch every step โ effort levels โ 5 settings from low to max. tune speed vs intelligence depending on the task...no more fixed thinking budgets โ verification โ the biggest one. give claude a way to test its own work. browser, bash, or computer use. boris ends every prompt with /go - a skill that tests, simplifies, and opens a PR opus 4.7 gives you 2-3x more output...but only if you verify most people are still using claude like it's 2024 github โ https://t.co/zCNrdWbgpv
@sharbel ยท
How To Use Claude Code Like a CTO: 0:00 Most people use Claude Code wrong 0:52 Why the terminal scares people away 1:21 What Claude Code actually is 2:49 Using Claude Code like a CTO 4:03 Feature 1, Claude.md 5:52 Feature 2, plan mode 6:35 Feature 3, skills and slash commands 7:20 Feature 4, hooks 7:59 Feature 5, MCP servers 8:37 Feature 6, subagents 9:20 Feature 7, ultra think 10:11 Bonus, Claude Code memory 10:49 My exact Claude Code workflow 13:32 The limitations 15:46 Real example, building ViralPen 17:02 Final takeaway
Watch video
@dbreunig ยท
My takeaways from scanning the Claude Code code for ~45 min this evening: 1๏ธโฃHarness engineering is hard. There's a lot of hard won knowledge in here and plenty of diagnostics to keep the feedback flowing. 2๏ธโฃHarnesses and prompts smooth out model quirks. @SrihariSriraman and I covered this last month, but good to see it verified here. So many conditionals based on model types and specific contexts to deploy to mitigate model weirdness. 3๏ธโฃSo much of this is CLI app boilerplate. Fully expect a tool like @badlogicgames's pi to be the foundation for any CLI agent being built today. I talk about the last point, the opportunity for shared foundations, in a post today: https://t.co/F9OFdOahYZ
@Suryanshti777 ยท
Most people use Claude Code like a chat. Power users use it like an operating system. This cheatsheet is basically the difference. Claude Code isnโt just prompting โ itโs sessions, agents, skills, hooks, MCP servers, and workflows. You can: โข spin isolated worktrees โข run multi-agent code reviews โข schedule cron-like loops โข batch parallel changes โข auto-load skills on trigger โข connect GitHub, Slack, DBs via MCP โข define project rules with CLAUDE.md โข switch reasoning effort on demand This is not a chatbot. This is programmable development infrastructure. The real power stack: CLAUDE.md โ agent behavior Skills โ reusable capabilities Subagents โ isolated workers Hooks โ automated actions MCP โ external tools & data Together = fully autonomous coding system. Most devs only use: โClaude, write this function.โ Power users use: /plan โ /review โ /batch โ /loop โ /debug And ship in hours, not days. Pro move: Set low effort for quick edits High effort for architecture Batch for large refactors Loop for monitoring tasks Once you learn the commands, Claude stops being an assistant and becomes a coding command center. Bookmark this. You wonโt remember 60+ commands when you actually need them.
@carlvellotti ยท
Fine, I'll admit it: I've wasted many hours optimizing parts of my Claude Code setup I never even used! Whole weekends writing a perfect CLAUDE files, setting up folder structures, building out the skills. Then barely using any of it and eventually forget how it even works. I'm sure I'm not the only one ๐ Here's what I've learned about building one that actually works: 1๏ธโฃ Grow from use, not planning. Don't design your OS upfront. Use Claude for real work. Notice what you keep repeating. That's what goes in your CLAUDE.md. The system grows from friction, not imagination. 2๏ธโฃ Don't automate until you've done it manually like 3+ times. The first time you do something, just do it. Also, probably the second time. The third time, MAYBE put it in CLAUDE(.)md. The sixth, maybe build a skill. You need to understand the task before you can automate it well. And you need to make sure it's something it's worth building out. 3๏ธโฃ Your CLAUDE(.)md is opinions, not descriptions. "I prefer bullet points over paragraphs. Be direct, skip preamble. When estimating, I want ranges not single numbers." NOT: "I am a Senior PM at Acme Corp. I have 8 years of experience in B2B SaaS." Claude doesn't need your resume. It needs to know how you think. That other context can live elsewhere to pull in when you need 4๏ธโฃ Flat is better than nested MOST of the time. One level of project folders handles 90% of PM work. Don't build a deep hierarchy you'll never maintain. 5๏ธโฃ Every project gets its own folder. The folder IS the context boundary. Outputs live with the work that produced them. When a project is done, the folder is the archive. Boom, so simple and it doesn't need to be more complicated than that. 6๏ธโฃ Few things ever need to reach full automation. Prioritize like a PM. Don't invest in building things out until you are 100% sure it's worth it. As every PM knows, building something is the beginning of the journey. Maintaining and improving has way more hidden costs than you can anticipate. Your OS should look like the work you actually do, not the system you wish you had. โโ Hi I'm Carl ๐ Want to learn more about Claude Code? I've got a 100% free courses where you learn Claude Code IN Claude Code. Over 25,000+ people have done them to rave reviews. โ PM version: https://t.co/QrU0bspGur โ General version: https://t.co/iTPut74HUs And if you want to take your skills even further, the application for Claude Code for PMs: Mastery is currently open! Advanced Claude Code lessons + active Slack community of PM Claude Coders + weekly office hours with me. Learn more and apply here: https://t.co/ASVkVmfsTY
@maheshnani122 ยท
How to use Claude Code effectively - CLAUDE. md: Define project rules (Claude reads this first) - Plan Mode: Ask for plan BEFORE coding - TDD Loop: Start with failing tests, go green - Git Worktrees: Parallel feature development - /compact: Keep context clean - /cost: Track token usage - Subagents: Run tasks in parallel - Skills: Reusable workflows for teams
@coreyhainesco ยท
Claude Code pro tip ๐ก Have Claude ask Codex to review its plan and implementation. Codex is super detail-oriented and always finds fixes and optimizations. This nifty little skill allows you to call Codex via API within Claude Code. It's become one of my go-to's every day. https://t.co/B4EIuPM4lB
@alexxubyte ยท
If Claude Code is a burger... Before each model call, Claude Code assembles a context window from 9 distinct sources. Think of it as a burger, each layer adds something different. 1. System Prompt: Defines Claude's role, behavior, and tone. This sets the foundation. 2. Environment Info: Git status, branch info, and current date. Pulled in via getSystemContext() 3. CLAUDE.md: A four-level instruction hierarchy: managed โ user โ project โ local. Plain-text Markdown, so users can read, edit, and version-control everything the model sees. 4. Auto Memory: Contextually relevant memory entries prefetched asynchronously. An LLM scans memory-file headers and surfaces up to 5 relevant files on demand. 5. Path-scoped Rules: Conditional rules that load lazily when the agent reads files 6. Tool Metadata: Skill descriptions, MCP tool names, and deferred tool definitions. 7. Conversation History: Carried forward across iterations. 8. Tool Results: File reads, command outputs, and subagent summaries. 9. Compact Summaries: When history grows too long, older segments are replaced by model-generated summaries.
@Shruti_0810 ยท
99% of people using Claude Code are barely scratching the surface. They open the terminal. Ask for a feature. Hope for the best. Meanwhile power users are running entire engineering workflows from prompts alone. So I turned dozens of scattered docs, Reddit threads, and workflows into ONE visual cheat sheet for Claude Code. It covers: โข hidden CLI flags โข session + memory management โข debugging workflows โข automation loops โข multi-agent setups โข review + lint flows โข MCP + tool integrations โข thinking controls โข monitoring + reporting โข keyboard shortcuts โข and productivity tricks almost nobody uses A few underrated gems from the sheet: โ /compact Compresses conversation history without losing context. Huge for long sessions. โ /review Makes Claude act like a senior engineer reviewing your codebase. โ /think Changes reasoning depth. Most people never touch it. Massive difference on hard tasks. โ /agents + /auto Lets Claude run repeated workflows almost like mini autonomous agents. โ /memory Persistent project instructions = dramatically better outputs over time. The crazy part: Most developers still use AI like a chatbot. The real unlock is treating Claude Code like an operating system for software development. Once you build proper workflows: โข PR reviews get faster โข debugging becomes systematic โข repetitive tasks disappear โข context switching drops hard โข and shipping speed goes insane Bookmark this. Youโll need it later. ๐
@Shruti_0810 ยท
Most people treat CLAUDE.md like a prompt file. Thatโs the mistake. If you want Claude Code to feel like a senior engineer living inside your repo, your project needs structure. Claude needs 4 things at all times: โข the why โ what the system does โข the map โ where things live โข the rules โ whatโs allowed / not allowed โข the workflows โ how work gets done I call this: The Anatomy of a Claude Code Project ๐ โโโโโโโโโโโโโโโ 1๏ธโฃ CLAUDE.md = Repo Memory (keep it short) This is the north star file. Not a knowledge dump. Just: โข Purpose (WHY) โข Repo map (WHAT) โข Rules + commands (HOW) If it gets too long, the model starts missing important context. โโโโโโโโโโโโโโโ 2๏ธโฃ .claude/skills/ = Reusable Expert Modes Stop rewriting instructions. Turn common workflows into skills: โข code review checklist โข refactor playbook โข release procedure โข debugging flow Result: Consistency across sessions and teammates. โโโโโโโโโโโโโโโ 3๏ธโฃ .claude/hooks/ = Guardrails Models forget. Hooks donโt. Use them for things that must be deterministic: โข run formatter after edits โข run tests on core changes โข block unsafe directories (auth, billing, migrations) โโโโโโโโโโโโโโโ 4๏ธโฃ docs/ = Progressive Context Donโt bloat prompts. Claude just needs to know where truth lives: โข architecture overview โข ADRs (engineering decisions) โข operational runbooks โโโโโโโโโโโโโโโ 5๏ธโฃ Local CLAUDE.md for risky modules Put small files near sharp edges: src/auth/CLAUDE.md src/persistence/CLAUDE.md infra/CLAUDE.md Now Claude sees the gotchas exactly when it works there. โโโโโโโโโโโโโโโ Prompting is temporary. Structure is permanent. When your repo is organized this way, Claude stops behaving like a chatbotโฆ โฆand starts acting like a project-native engineer.
@Hartdrawss ยท
you should 100% be claudemaxxing i've tried every plugin, wrapper, and github repo promising to give my agent superpowers cursor rules generators, custom mcp stacks, orchestration layers that do things without even telling you whats happening under the hood honestly? i cant tell if any of them moved the needle reason is simple. you get better at something the more you use it. every abstraction layer you add is a thing you stop understanding AI labs are shipping at insane speed. if you're missing a feature, wait 3 weeks. it'll ship. so here's my actual claude setup. no plugins. no wrappers: 1/ cursor for file-level edits. claude code for full feature builds. they do different things, stop using them the same way 2/ supabase MCP connected directly. it reads and writes the db without you copy-pasting schema into every prompt 3/ chrome devtools MCP connected. claude can inspect the dom, read console errors, and debug frontend issues without you relaying anything 4/ write a CLAUDE[.]md file at the root of every project. it loads context automatically every session. no re-explaining your stack, conventions, or rules ever again 5/ ask claude code to write your cursor rules file. let the AI configure the AI. sounds stupid, works every time 6/ use sub-agents for long tasks. one agent plans, another executes. context stays clean and outputs get sharper end of day: ask it to write the git commit message and update the docs. 10 minutes of work you will never do manually again been averaging 10-20% faster task completion since i stopped trying to hack the tools and just learned them Now go claudemaxx !
@Hartdrawss ยท
if you open Claude Code without a structured workflow, you probably hate money. the skill gap isn't knowing prompts. it's knowing which command to run before you touch the terminal. here's the exact workflow I used from @mattpocockuk 1. start with `/grill-me` - paste your app idea or plan - Claude will ask you 16 to 50 questions before it does anything - mine ran 38 the first time i tried it - it walks every branch of the decision tree, resolving dependencies one by one - you fix the broken assumptions before they become broken code 2. move to `/to-prd` - converts the grilling conversation into a proper requirements doc - skips the steps you already covered - doesn't start from scratch - outputs user stories, not implementation notes - lands as a GitHub issue with a triage label - normal team workflow, no AI sidetrack 3. then `/to-issues` - reads the PRD and breaks it into independently-grabbable vertical slices - each issue is tagged HITL (you stay in the loop) or AFK (agent executes solo) - dependency-sorted so nothing blocks anything 4. finally `/tdd` - now the agent writes code. red-green-refactor - can't start green if red hasn't failed - phase-gated. no shortcuts. Hope this helps !
@socialwithaayan ยท
you can fix claude code's worst habits with a single markdown file, and it's free ๐คฏ andrej karpathy, co-founder of openai and former ai director at tesla, posted his own field notes after going from 80% manual coding to 80% agent coding in a couple months. the pattern he named: models make wrong assumptions and just run with them, they overcomplicate simple fixes, they quietly touch code they don't understand. someone took those exact observations and turned them into four enforceable rules for claude code. why it works: instead of a generic "please write clean code" instruction, it gives claude four specific behaviors to check itself against, each one mapped directly to a failure mode karpathy actually described. what you get for $0: โ one CLAUDE. md file with four principles, drop it in and go โ think before coding: state assumptions out loud, ask instead of guessing โ simplicity first: if 200 lines could be 50, rewrite it โ surgical changes: every edited line traces back to the actual request โ goal driven execution: "fix the bug" becomes "write a failing test, then make it pass" โ also ships a cursor rule for the same guardrails outside claude code what it replaces: โ ai quietly bloating a 50 line fix into 300 โ silent guesses on ambiguous requests instead of a quick clarifying question โ unrelated "cleanup" edits nobody asked for how to set it up (2 minutes): 1. copy the four principles from the readme 2. save as CLAUDE .md in your project root, or append to an existing one 3. or install as a plugin: /plugin install andrej-karpathy-skills@karpathy-skills 4. it's biased toward caution, so use judgment on trivial one-liners the honest part: this specific repo is brand new, but the idea behind it isn't. karpathy's notes are from january, and other developers built claude. md files off the same observations back in april, some hitting big numbers of their own. this is just the latest fork of a concept that already had its viral moment. 161 stars, MIT license, free and open source.
@alphabatcher ยท
Claude Code creator gave a 24-minute prompting workshop "Anthropic technical onboarding used to take 2-3 weeks" Boris says Claude Code Q&A brought it down to about 2-3 days The order he teaches: 1. Start with codebase Q&A Before edits, make Claude explain the repo Ask: - where is auth handled - how is this class used - how do I instantiate this object - why does this function have 15 arguments - look through Git history and explain why this changed Claude reads files, follows examples, checks Git history, follows linked issues, and explains why the code exists 2. Make it plan before it edits Use this prompt: "Before you write code, make a plan Brainstorm options Name the files you need Ask for approval before changing anything" Claude should explore first, plan second, edit third 3. Give it a way to check its work A coding agent improves when it can see the result Give it: - unit tests - integration tests - Puppeteer screenshots - iOS simulator screenshots - lint and typecheck commands Prompt it like this: "Build this UI Run it in the browser Take a screenshot Compare it to the mock Fix the differences Stop after 3 passes" 4. Put repo context in `CLAUDE.md` Boris says `CLAUDE.md` is loaded at the start of every session Put in: - test commands - build commands - style guide - common MCP tools - important files - architecture decisions - repo-specific gotchas Keep it short Long context burns the session before the work starts 5. Split shared and personal context Shared repo rules go in project `CLAUDE.md` Personal preferences stay local Nested `CLAUDE.md` files can live inside child folders, so Claude pulls in folder-specific rules only when it works there 6. Turn repeated work into slash commands Use `.claude/commands` for repo tasks you repeat Examples: - label GitHub issues - generate release notes - check a PR - run the project test sequence - prepare a weekly shipped summary 7. Use shortcuts while it runs - type `#` to make Claude remember a rule and fold it into `CLAUDE.md` - type `!` to run a shell command and add the output to context - hit `Esc` when an edit is almost right and one line needs changing - hit `Esc` twice to jump back in history - hit `Ctrl+R` to see the same output Claude sees 8. Use `claude -p` for scripts For CI, incident response, or recurring checks: - pass a prompt - allow specific tools - request JSON or streaming JSON - pipe into it and out of it like a Unix utility
Watch video
@smratitiwa86867 ยท
Most people think using Claude Code is about writing better prompts. Itโs not. The real unlock is structuring your repository so Claude can think like an engineer. If your repo is messy, Claude behaves like a chatbot. If your repo is structured, Claude behaves like a developer living inside your codebase. Your project only needs 4 things: โข the why โ what the system does โข the map โ where things live โข the rules โ whatโs allowed / forbidden โข the workflows โ how work gets done I call this: The Anatomy of a Claude Code Project ๐ โโโโโโโโโโโโโโโ 1๏ธโฃ CLAUDE.md = Repo Memory (Keep it Short) This file is the north star for Claude. Not a massive document. Just three things: โข Purpose โ why the system exists โข Repo map โ how the project is structured โข Rules + commands โ how Claude should operate If CLAUDE.md becomes too long, the model starts missing critical signals. Clarity beats size. โโโโโโโโโโโโโโโ 2๏ธโฃ .claude/skills/ = Reusable Expert Modes Stop repeating instructions in prompts. Turn common workflows into reusable skills. Examples: โข code review checklist โข refactoring playbook โข debugging workflow โข release procedures Now Claude can switch into specialized modes instantly. Result: More consistent outputs across sessions and teammates. โโโโโโโโโโโโโโโ 3๏ธโฃ .claude/hooks/ = Guardrails Models forget. Hooks donโt. Use hooks for things that must always happen automatically. Examples: โข run formatters after edits โข trigger tests after core changes โข block sensitive directories (auth, billing, migrations) Hooks turn AI workflows into reliable engineering systems. โโโโโโโโโโโโโโโ 4๏ธโฃ docs/ = Progressive Context Donโt overload prompts with information. Instead, let Claude navigate your documentation. Examples: โข architecture overview โข ADRs (engineering decisions) โข operational runbooks Claude doesnโt need everything in memory. It just needs to know where truth lives. โโโโโโโโโโโโโโโ 5๏ธโฃ Local CLAUDE.md for Critical Modules Some areas of your system have hidden complexity. Add local context files there. Example: src/auth/CLAUDE.md src/persistence/CLAUDE.md infra/CLAUDE.md Now Claude understands the danger zones exactly when it works in them. This dramatically reduces mistakes. โโโโโโโโโโโโโโโ Hereโs the shift most people miss: Prompting is temporary. Structure is permanent. Once your repository is designed for AI: Claude stops acting like a chatbot... โฆand starts behaving like a project-native engineer. ๐
@zaimiri ยท
the guy who built Claude Code runs a 100-line CLAUDE.md you don't need 800. boris cherny is a staff engineer at anthropic. he created the tool. he runs 10-15 sessions in parallel daily. his file is ~2,500 tokens. the 6 rules he actually uses: > plan mode before every task > subagents for parallel work > verify before calling anything done > demand elegance, no hacks > fix bugs autonomously without asking > self-improvement loop that last one is the real weapon. every time claude makes a mistake, they add a rule. his exact instruction: "after every correction, end with: Update your CLAUDE.md so you don't make that mistake again." the ai gets better on YOUR codebase. not just in general. specifically yours. he hasn't written a single line of SQL in 6+ months. claude pulls bigquery data directly via cli. everyone's building custom frameworks. the creator uses a text file. key lesson: keep it under 200 lines.
@Layton_Gott ยท
I don't write code... I describe what feature I want and Claude Code builds it. Here's my exact process I use at 17: 1. I describe the feature in plain English. Just what it should do and why. The clearer I am here, the better everything after this goes. 2. I give Claude Code the full context. It reads my codebase, understands my file structure, my naming conventions, my existing patterns. Remember context is everything. 3. Claude Code proposes a plan before writing anything. Which files it will create, which ones it will edit, and why. I review this before a single line gets written. 4. I answer its questions. It asks about ambiguous decisions instead of guessing. One wrong assumption early ruins everything downstream. This step saves more time than any other. 5. Claude Code executes. Creates the branch, writes across multiple files, keeps consistency with my existing codebase. I watch in real time and course-correct if needed. 6. I review the output. Not line by line. Check the logic, test the feature, make sure it works. Most of the time it's 90% done. I clean up the last 10% myself. 7. I deploy it. This is how I build every feature for Scribe daily. (bookmark this)
@JustAnotherPM ยท
Two Anthropic engineers who built Claude Code sat down and showed every feature most users have never touched. Bookmark this. Come back to it this weekend. Here is what they covered: ๐ญ. ๐ ๐๐น๐๐ถ-๐ณ๐ถ๐น๐ฒ ๐ฒ๐ฑ๐ถ๐๐ถ๐ป๐ด. Claude Code rewrites across your entire codebase in one pass. Not file by file. ๐ฎ. ๐ฆ๐๐ฏ๐ฎ๐ด๐ฒ๐ป๐๐. Spawn background agents that research, test, or review while you keep working in the main session. ๐ฏ. ๐๐ผ๐ผ๐ธ๐. Set guardrails that run before or after every tool call. Block destructive commands automatically. ๐ฐ. ๐ ๐๐ฃ ๐ถ๐ป๐๐ฒ๐ด๐ฟ๐ฎ๐๐ถ๐ผ๐ป๐. Connect Claude Code to Slack, GitHub, databases, and any API with one config file. ๐ฑ. ๐๐๐๐จ๐๐.๐บ๐ฑ. The file that turns a generic model into a teammate who knows your project, your standards, and your shortcuts. 24 minutes. Free. From the people who built the tool.
@_vmlops ยท
THIS SINGLE FILE WILL CHANGE HOW YOU USE CLAUDE CODE FOREVER it's called andrej-karpathy-skills karpathy publicly called out exactly how llms fail at coding: silent assumptions, bloated code, touching things they shouldn't forrestchang took that and turned it into 4 rules inside one CLAUDE.md file โซ๏ธ ask don't assume โซ๏ธ write 50 lines not 500 โซ๏ธ touch only what you must โซ๏ธ give it a finish line not a to-do list that last one rewires everything instead of "fix the bug" try "write a test that reproduces it, then make it pass" claude loops until it hits the goal. zero hand-holding needed drop it into any project in seconds: curl -o CLAUDE.md https://t.co/xKd8X2uHWC
@aaliya_va ยท
You can now fix Claude Code's bad habits with a single command. Claude Code is powerful but has a habit of repeating the same mistakes. It edits files in loops, ignores corrections, and drifts from your original request. A new open-source CLI tool called claude-doctor fixes this. It reads your local session transcripts and finds behavioral anti-patterns automatically. It tracks 11 signals across three categories: 1. Structural: edit-thrashing, error loops, excessive file reads 2. Behavioral: ignored corrections, repeated instructions, negative drift 3. Lexical: sentiment scoring with agent-specific tokens The tool then generates rules you paste into your CLAUDE.md. These rules tell the agent how to stop repeating its worst habits, based on your actual history. One command analyzes 838+ sessions. It detects when the agent edits the same file 5+ times, fails 3 tools in a row, or ignores your "no." Run npx claude-doctor and let your transcripts do the coaching. Link in comments. โ
Watch video
@CodeByPoonam ยท
90% of people are using Claude Code wrong. Anthropic's internal workflow is just 4 steps: โ Explore (Plan Mode, zero changes) โ Plan (detailed implementation plan before any code) โ Implement (with tests so Claude verifies itself) โ Commit (let Claude own git too) Most people skip the first two and wonder why results are inconsistent.
@RoundtableSpace ยท
Most developers use Claude Code sequentially, but configuring Opus 4.8 with the ultracode effort mode turns your terminal into a fully autonomous multi-agent engineering workflow. Here is how the automated workflow operates under ultracode: - Run /effort ultracode or set the model to Opus 4.8 with xhigh reasoning effort - Claude analyzes your prompt and automatically assesses whether the task requires parallel execution - The system constructs dynamic orchestration scripts to spin up specialized subagents without manual setup - Subagents execute research, planning, implementation, and code review concurrently across your repository Instead of handling tasks step-by-step, you give a high-level objective and let Claude manage the end-to-end development loop.
Watch video
@nrqa__ ยท
Someone turned Claude Code into a full game dev studio. One repo. 48 AI agents working together. It's called Claude Code Game Studios. You open Claude Code, type /start, and suddenly you have a creative director, a technical director, a producer, department leads, and 40+ specialists ready to ship your game. Not one assistant trying to do everything. A real studio hierarchy. Directors guard the vision. Leads own their domains. Specialists do the hands-on work. Each one knows when to escalate and when to shut up. Here's what's inside: โ 48 specialized agents across design, programming, art, audio, narrative, QA, and production โ 37 slash commands. /sprint-plan, /code-review, /balance-check, /brainstorm, /playtest-report, /hotfix, and more โ 8 hooks that run automatically. Block hardcoded values on commit. Warn on force push. Validate asset naming. Log every agent call. โ 11 path-scoped rules. Edit gameplay code and it enforces delta time usage. Edit networking and it enforces server-authoritative patterns. Automatic. โ 29 document templates. GDDs, ADRs, sprint plans, economy models, faction design, all ready to fill. Three-tier hierarchy mirroring real studios: Tier 1 runs on Opus. Creative director, technical director, producer. The vision guards. Tier 2 runs on Sonnet. Game designer, lead programmer, art director, audio director, narrative director, QA lead. Department owners. Tier 3 is the specialist floor. Gameplay programmer. AI programmer. Network programmer. Level designer. Economy designer. Technical artist. Sound designer. UX designer. Accessibility specialist. Live ops. Community manager. Plus dedicated engine specialists. Pick one. โ Godot 4 (GDScript, Shaders, GDExtension) โ Unity (DOTS/ECS, Shaders/VFX, Addressables, UI Toolkit) โ Unreal Engine 5 (GAS, Blueprints, Replication, UMG/CommonUI) Here's what makes it actually usable: It's not autopilot. Every agent follows the same protocol. Ask questions first. Present 2-4 options with pros and cons. Wait for your call. Draft before finalizing. Never ship without sign-off. You stay the boss. The agents just make sure nothing slips through. The hooks are the real unlock. Commit code with a magic number and it gets flagged. Open a session on a fresh project and it suggests /start. Push to main and it warns you. Spawn a subagent and it logs the trail. All automatic. Nothing to configure. Grounded in real game dev frameworks too. MDA. Self-Determination Theory. Flow state design. Bartle player types. This wasn't vibe-coded. The setup is 3 steps: โ Clone the repo as your project template โ Open Claude Code โ Run /start and it asks where you are (no idea, vague concept, clear design, existing work) and routes you to the right workflow That's it. You now have a studio. Solo devs spend years learning what a producer does. Months figuring out when to write a GDD vs prototype. Weeks debating architecture with nobody. This gives you all of it on day one. 15.3K stars. 2.2K forks. Active development. MIT License. 100% Open Source.
@jbarbier ยท
Claude Code is much more than a coding tool on steroids. It has a profound impact on how we think about tackling problems, challenges, and makes our lives much more fun as a result. I built a @davidgoggins fitness coach on WhatsApp for my friends with Claude Code. It's unhinged. It's motivating. We love it. I've shipped 12 apps like this since January, all tailored to me and my friends' and my family's actual lives. Multiply this by millions of people. Society is being reshaped right now. Stay hard. ๐ซต
@subinium ยท
My personal @claudeai code patterns: โข Set effort to max in /model before starting โข Use ultrathink + ultraplan liberally โข If you have Codex โ use /codex plugin for code-level review before commits (UI/content checks stay in cc) โข Don't fill the 1M context. compact or handoff around 20โ40% โข Periodically review /insights and update your claude.md accordingly โข For new toy projects: rough UI draft โ extract BI โ redesign based on it โข For open source: define design principles early, then expand from there โข Don't bloat your skills. The most useful ones I added were per-project CI process skills โข Harnesses aren't magic. Unless you're going deep on a vertical or betting on open source, don't over-invest in setup. For general use, pick and choose from everything claude code, omo, omc, omx
@0xDvnl ยท
Most people treat Claude Code like a chatbot that writes code. Here's how I actually set it up: 1. Give it account-level access to Vercel and Supabase. No more copy-pasting queries and deployment configs back and forth. 2. Spin up a sub-agent just for QA. It reviews everything the main agent builds. Catches things before you do. 3. Connect it to https://t.co/IVgekVmkXe. Your coding agent can now request custom-generated images, icons, illustrations โ and implement them directly. No more emoji placeholders. No more generic stock assets. 3x perceived UI quality overnight. The principle: think WITH Claude, but empower it to build nearly autonomously. You're the architect. It should have its own tools.
Best Claude Code tweets
Xholic studies what works in your niche, drafts posts in your voice and schedules them for the hours your audience is online.
No card needed ยท Nothing to cancel
Browse all tweet collectionsLive demo: no sign-up or login needed ยท takes about 2 minutes
Keep exploring