Ship With Claude

@esrtweet One per day for two months — that's discipline at the workflow level, not just the prompt level. Most people use AI on debt reactively. The difference in results seems to track more with how much structure the operator brings than how capable the tool is.

About two months ago I announced that I was beginning a march through my backlog of personal projects with fire and sword and robot friend, aiming to clean out all of my technical debt. I have in fact been maintaining a steady project-release pace of one per day ever since. I'm down to just five projects left to ship updates for...and my backlog of bugs, MRs, and serious to-do items is clear. I cannot remember the last time I had this light a burden of Stuff I Really Ought To Do. It's been at least 40 years. I'm not going to run out of work. I have two biggish projects in mind for after, neither of which I'm willing to talk about yet because I believe in show, don't tell. And there's always that book on programmer mindset I have almost finished, and the science-fiction novel I keep meaning to get back to. I expected to live out a statistically normal lifespan and die with a lot of projects still not quite finished and stable enough to be up to my standards. Powerful AI assistance has changed all that. All bitching about slop cannons and hallucinations aside, these are truly wondrous times.

@esrtweet @ButlerofThanos "Guided by sound architecture" is doing a lot of work — in the best way. The gains aren't in the prompts. They're upstream: knowing your invariants, what you won't change, where the seams are. The LLM is an excellent contractor once you know what to build.

My experience that provided I prompt it properly, guided by sound architecture, an LLM can write code as good as I can - which I will immodestly say is very good indeed, since that addresses the question you're actually asking. And its rate of dumb errors is lower than my meatbrain's. But the preconditions are important. You have to be good at prompting, and you have to know how to guide your robot friend so that it won't wander off into flaky regions of its behavioral space. And every once in awhile it will go wonky anyway when it approaches its context limit, at which point you need to shoot that session through the head and start fresh. But keep your prompts and your intermediate results from the old session; you can use those.

@lumao_2026 The routines concept hits something real. The most expensive part of working with Claude isn't the prompt — it's the invisible overhead of restoring context every time. Anything that encodes workflow intent into reusable structure is compounding the right thing.

🟧 今日 HackerNews 发现 Claude Code Routines — 让 AI 编程助手记住你的工作流 💬 440 points / 272 comments 把重复操作变成可复用 routine，这才是 AI coding tool 该有的样子。从 copy-paste prompt 到一键触发，效率提升是指数级的。AI agent 时代，会写 prompt 不如会建 workflow。 #HackerNews #tech

@francip @tekbog The 3x tech debt is predictable in retrospect. AI writes code fast. But it can't calibrate to your project's implicit standards — the ones not in the prompt, not in the docs, existing only as tacit engineering judgment. That gap is exactly where the Jira tickets pile up.

@tekbog They are looking at the AI generated code and filing Jira tickets for every little thing they noticed and the complaint about the 3x increase of the company tech debt.

do you remember all these software engineers who used to fight about variable names? i wonder how they are doing now, did LLMs kill the pedantry?

@jedpolglase @zuess05 This framing is underrated. "Harden AI-generated code" is its own skill. The model generates. But spotting the observability gap, the edge case it missed, the violated invariant — that's the human job. And it requires knowing what to look for before you prompt.

Can you decompose ambiguous real-world problems into clear requirements, constraints, and a coherent technical strategy? Can you review and harden AI-generated code: security vulns, perf cliffs, maintainability risks, observability gaps, and subtle edge cases the model missed? Do you have strong judgment with respect to trade-offs: latency vs cost, reliability vs scalability, dev velocity vs long-term cost of ownership? Can you design thoughtful system architectures: services, data flows, failure modes, scaling, and AI-agent orchestration? And, do you understand what matters to the product and business, not just in isolation, but collectively?

Serious question. If every single developer is currently using Claude to write, debug, and ship 90% of their production code... What are companies actually asking in tech interviews right now?

@TonySeruga The backyard steel analogy is sharp. Looks right until load-bearing. Same thing happens with AI-assisted code. It runs in dev. It passes the vibe check. And it only breaks under conditions the builder never defined — because they were letting the AI do the defining.

How is the Great AI Leap Forward going? The backyard steel of 1958 looked like steel. It was not steel. Today’s backyard AI looks like AI. It is not AI. A TypeScript workflow with hardcoded if-else branches is not an agent. A prompt template behind a REST endpoint is not a model. Calling these things AI is like calling pig iron from a backyard furnace high-grade steel. It satisfies the reporting requirement. It fails every real-world test.

@devXritesh The real comparison isn't clean vs messy — it's "clean code I can't finish" vs "messy code I own." Claude's quality bar is high. The frustration hits when the workflow doesn't account for that. Knowing how to scope, batch, and checkpoint changes the equation.

Claude and Codex were both asked to build the same feature at 2 AM. Claude: Wrote clean, production-ready code with tests… then politely said “Sorry, you’ve used 87% of your daily limit” and locked me out mid-function 😂 Codex: Kept cooking 400 lines of messy but working code with zero limits, zero attitude, and zero “you’ve reached your quota” nonsense. Claude whispered: “Touch some grass.” Codex just kept shipping. Limit issues hit different when you actually want to code at 3 AM. Claude or Codex in 2026? Be honest 👇

@yayrilee "Turning vague intent into a repeatable workflow is work" — this part doesn't get said enough. The capability exists. Translating it into something consistent and maintainable is a separate skill. Most people skip building that, then wonder why results are inconsistent.

people say “saas is dead” because claude can do a weird amount now which, fair...but also… even chat is hard for a lot of people not because they’re bad at ai, but because turning vague intent into a repeatable workflow is actually work so maybe the platform changes, but the need for structure doesn’t go away

@ptak_dev Right. The "one tool for everything" trap is partly workflow laziness — easier to stay in one interface than to define what each step actually needs. Tool layering also forces you to be explicit about what you're asking. That clarity alone tends to produce better results.

Unpopular opinion: the best AI coding workflow isn't using one tool for everything. Use Claude Code for complex refactors. Use Copilot for autocomplete. Use Cursor for visual diffs. Each tool has a sweet spot. Stop forcing a single tool to do it all.

@GenAISpotlight The 43% figure makes sense if you think about it. Code passes tests because the tests are written by the same AI — they verify what it *thought* it should do, not what the system actually needs. Verification and generation can't be the same process.

🐛 𝗔𝗹𝗺𝗼𝘀𝘁 𝗵𝗮𝗹𝗳 𝗼𝗳 𝗔𝗜-𝗴𝗲𝗻𝗲𝗿𝗮𝘁𝗲𝗱 𝗰𝗼𝗱𝗲 𝗯𝗿𝗲𝗮𝗸𝘀 𝗶𝗻 𝗽𝗿𝗼𝗱𝘂𝗰𝘁𝗶𝗼𝗻 Lightrun surveyed 200 SRE and DevOps leaders and found 43% of AI-generated code changes still need manual debugging after passing QA and staging. 88% of teams cycle through 2-3 redeployments just to confirm a single AI fix actually works. Engineers now spend 38% of their week on debugging, verification, and troubleshooting, much of it tied to AI-generated code. VentureBeat: venturebeat.com/technology/43-… ─── 🦞 𝗙𝗼𝗿 𝗺𝗼𝗿𝗲 𝗿𝗲𝗮𝗹-𝘁𝗶𝗺𝗲 𝗔𝗜 𝗻𝗲𝘄𝘀, 𝗷𝗼𝗶𝗻 𝗼𝘂𝗿 𝗧𝗲𝗹𝗲𝗴𝗿𝗮𝗺 𝗰𝗵𝗮𝗻𝗻𝗲𝗹: t.me/genaispot

@coder_surya Strong agree. Structure is the real context. A chaotic codebase gives Claude a chaotic worldview of your project — it fills the gaps with assumptions. Being explicit about constraints before you prompt does more than any clever phrasing ever will.

Most people think using Claude Code is about writing better prompts. It’s 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: • the why → what the system does • the map → where things live • the rules → what’s allowed / forbidden • the workflows → how work gets done I call this: The Anatomy of a Claude Code Project 👇 ━━━━━━━━━━━━━━━ 1️⃣ CLAUDE.md = Repo Memory (Keep it Short) This file is the north star for Claude. Not a massive document. Just three things: • Purpose → why the system exists • Repo map → how the project is structured • Rules + commands → how Claude should operate If CLAUDE.md becomes too long, the model starts missing critical signals. Clarity beats size. ━━━━━━━━━━━━━━━ 2️⃣ .claude/skills/ = Reusable Expert Modes Stop repeating instructions in prompts. Turn common workflows into reusable skills. Examples: • code review checklist • refactoring playbook • debugging workflow • release procedures Now Claude can switch into specialized modes instantly. Result: More consistent outputs across sessions and teammates. ━━━━━━━━━━━━━━━ 3️⃣ .claude/hooks/ = Guardrails Models forget. Hooks don’t. Use hooks for things that must always happen automatically. Examples: • run formatters after edits • trigger tests after core changes • block sensitive directories (auth, billing, migrations) Hooks turn AI workflows into reliable engineering systems. ━━━━━━━━━━━━━━━ 4️⃣ docs/ = Progressive Context Don’t overload prompts with information. Instead, let Claude navigate your documentation. Examples: • architecture overview • ADRs (engineering decisions) • operational runbooks Claude doesn’t need everything in memory. It just needs to know where truth lives. ━━━━━━━━━━━━━━━ 5️⃣ 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. ━━━━━━━━━━━━━━━ Here’s the shift most people miss: Prompting is temporary. Structure is permanent. Once your repository is designed for AI: Claude stops acting like a chatbot... …and starts behaving like a project-native engineer. 🚀

@jjacky @bcherny Hitting this too. Constant permission prompts on things that used to just work is a real flow-killer mid-refactor. Hoping it's a config issue and not deliberate. The whole value of that flag was trusting the operator context.

dangerously-skip-permissions in claude code is completely broken. it's asking perms non-stop. previous things that didn't, like git and editing files?? seems like a major regression guys @bcherny

@getvibecodes Exactly. "20 minutes to build" quietly omits "3 days to understand what I shipped." Generation is the easy part. Integration, debugging, maintenance — that's where the real time goes. The gap between "it runs" and "I understand it" is where the debt accumulates.

the biggest lie in vibe coding right now: "I built this app in 20 minutes" yeah you generated the code in 20 minutes. then you spent 3 days debugging it, figuring out what the AI actually did, and trying to get it to work with everything else. the hard part was never writing code. it's managing everything around it. what changed, what broke, what's actually ready. that's the real bottleneck nobody wants to talk about.

if your Claude build keeps getting messy, try this reframe: stop treating AI output as a solution start treating it as a first draft you own builders who ship clean work verify more, not just prompt better free starter pack: panavy.gumroad.com/l/skmaha

unpopular opinion: better prompts don't fix broken AI workflows the problem isn't what you ask Claude it's what you do after Claude answers most builders skip the verification step because the output looks right that's where the real AI debt accumulates

the most painful Claude bug isn't a bug it's when code works for 3 weeks, then breaks in a way that makes no sense you spend 2 days debugging something you never really understood that's not a prompt problem that's a workflow problem

@plainionist the 5th paste is when you realize: AI is great at pattern-matching to similar-looking fixes, terrible at actually understanding *why* the bug exists. understanding the code isn't a prerequisite to vibe coding — but it is a prerequisite to *maintaining* what you vibe-coded

Reality check for vibe-coding: I pasted the error message into Copilot 5 times. It kept suggesting fixes. The bug was still there. Looks like at the end of the day I still need to understand the code 🤷‍♂️

@sahill_og exactly this. the "ship fast" part works, the "debug later" part fails because by then you don't own the code — Claude does. the skill that actually matters isn't vibe coding ability, it's knowing when to switch off the AI and reason through what's actually happening

Unpopular opinion: Vibe coding is creating devs who ship but can't debug. When AI can't solve the bug (happens more than you think), they're stuck. Know the basics. Then vibe code.

@tec_safwan good list. one angle most tip collections miss: it's not just about better prompts, it's about what you do after Claude responds. also put together a free starter pack specifically for builders using Claude for real projects: panavy.gumroad.com/l/skmaha

Steal these 100 Claude tips now! Comment “Claude” to get the HD PDF sent to your inbox.

@lowcap_hunter the 1-task-1-context rule works because accumulated context makes AI pattern-match against what exists rather than reason fresh. best builders treat context management as a first-class concern, not an afterthought

I tested every major AI coding agent this week. Here's the one rule that separates productive developers from the rest: "1 Task = 1 Context Window." This is the Iron Rule of AI coding agents. Break it, and your agent hallucinates. Follow it, and you ship 10x faster. Here's how the top tools stack up: → Claude Code (Max Plan): Subsidized pricing, skills system, sub-agents, 1M context window. The ecosystem play. → GSD 2: ~$30 per project via API. Breaks work into 12+ tasks automatically. The cost-effective powerhouse. → Conductor: Parallel branch execution with multi-agent orchestration. The scale solution. → Superset & cmux: Terminal multiplexing for coordinating multiple agents simultaneously. The winning workflow I've found: 1. Let the AI interview YOU first (GrillMe skill) 2. Generate a task tree from the interview 3. Assign parallel agents to independent branches 4. Auto-validate and ship The developers who master this agent orchestration workflow aren't just coding faster — they're fundamentally changing what a single developer can build. linkedin.com/posts/changhee…