AI Coding in Practice

I Built a Fully Automated Short-Video Pipeline with Claude Code. Two Months In, Here's the Real Data

Sat, 18 Jul 2026 00:00:00 GMT

Script to finished video to scheduled publishing, zero human hands, daily posts on two platforms across three accounts. This is not a 'six figures a month' post — most of my videos got two-digit view counts. But three things the data taught me are worth more than a hit would have been: topic tier, content-platform fit, and the ceiling baked into the format. Full architecture plus 9 things that broke.

Claude Code or Codex? You're Asking the Wrong Question

Tue, 16 Jun 2026 00:00:00 GMT

Every few days someone on a dev forum asks 'is Claude Code or Codex better right now', and the thread splits into two camps that never convince each other. My take: it never resolves because the question is wrong. They aren't an either/or—they're two tools with different feel. The pragmatic move is to keep both, switch by task and quota, and let cost decide your daily driver. And the ultimate selection criterion was always 'which one finishes the job within your limit'—which requires getting the bill straight first.

AI Coding Kills Middle Managers First, Not Juniors or Seniors

Sun, 31 May 2026 00:00:00 GMT

Meta, Google, Amazon all flattened middle management before AI Coding hit scale; AI just turned that move from optional to mandatory. Three TL workflows — task allocation, review, cross-team alignment — get bypassed by senior ICs with agents, drowned by PR throughput, replaced by shared specs. Two paths left: collapse back to IC, or expand span upward.

Next Month, Big Tech Starts Auditing the Token Bill

Sat, 30 May 2026 00:00:00 GMT

For a year, companies rolled out AI coding tools on a 'just use it' basis—nobody watched tokens. But once it goes from 'a few people trying' to 'everyone, daily', the absolute bill clears finance's attention threshold. I bet cost auditing becomes a main thread within a month—and most teams can't answer the three questions (where did it go, how much was wasted, was it worth it) because vendor dashboards only show org-level totals. This is a tooling-category window.

Lagou Is Dead: A 13-Year Job Board Killed by AI Pulling Out All Three Middleman Steps at Once

Sat, 30 May 2026 00:00:00 GMT

Lagou — for most of the 2010s China's go-to job board for tech — didn't die at the tail end of a hiring winter. It's the marker event for AI's first replacement wave landing on middleman work: JD writing, resume screening and first-round interviews all got absorbed into corporate HR toolchains at once, leaving the platform with none of the three actions it lived on. Which middleman platform falls next? Three signals to watch, plus a sharper question: a lot of veterans are middlemen themselves and haven't noticed.

Why Vibe Coding Is Addictive: It Turns Coding Back Into a Game

Sat, 30 May 2026 00:00:00 GMT

Flow isn't mystical — it's just treating the task in front of you like a game. Games pull you into flow through three things: exploration, randomness, and a sense of achievement. Vibe coding happens to stuff all three back into writing software, which is exactly why it's addictive. But the same loop that delivers flow can also turn you into someone yanking a slot-machine lever. The veteran's job is to use the mechanic, not be used by it.

8-Line Spec Beats 8-Hour Prompts: The Real Payoff of Spec-Driven AI Coding

Fri, 29 May 2026 00:00:00 GMT

'I spent the whole afternoon arguing with the AI to make it write the right thing' — still a 2026 cliché. But teams that switched to writing 8-line specs upfront find AI's first-pass correctness jumps from 30% to 85%. This isn't a prompt-engineering trick. It's a software-engineering paradigm shift: the bottleneck has always been your inability to specify.

Building a Personal AI Coding Harness: 14 Traps Between "It Runs" and "It Works"

Mon, 25 May 2026 00:00:00 GMT

A harness isn't a pile of tools — it's the code that encodes your team's workflow as scheduling logic. A year of running one in production, sorted into 14 traps across 4 groups: economics (how you spend), robustness (how you don't crash), quality (how you ship), extensibility (how it evolves). Miss these and your harness stays a demo forever.

Experience Sediment 2.0: Stop Writing for Colleagues, Start Writing for Agents

Sun, 24 May 2026 00:00:00 GMT

Wikis, RFCs, and onboarding docs were written for "a future colleague." Since 2024 the most frequent reader of all that has stopped being human — it's an agent. The reader changed; most senior engineers haven't changed how they write. Here's the three-layer architecture (cursorrules / spec kits / harness orchestration) and which lessons are worth feeding an agent at all.

The Hardest Part of Rolling Out AI Coding Isn't Tech — It's Two Kinds of Old-Timers

Sun, 24 May 2026 00:00:00 GMT

A year of pushing AI Coding into a 20-person team. The blocker was never technology — it was two specific people. The mocker tries Cursor twice, gets burned, and tells everyone it doesn't work. The blocker is a mid-level leader who stalls forever behind 'quality risk' and 'compliance review.' Different motives, different counters. Mix the playbooks and you lose.

How to Become a Super Individual: Taste Is the Ticket, Tokens Are the Fuel, Brainpower Is the Ceiling

Sun, 24 May 2026 00:00:00 GMT

The hottest 2026 narrative is the 'super individual' or 'one-person company'. Strip away the story and look at the structure: three deeply unequal resources decide whether you can pull this off — taste that defines product shape, ruthless compression of token spend, and a brain that simply cannot keep pace with AI. The first two are the price of entry. The third is the actual ceiling.

The AI-Era Couple Business: She Builds Audience, He Builds Product

Thu, 21 May 2026 00:00:00 GMT

The strongest startup unit isn't a lone genius or a VC-backed team of ten — it's two people with deeply complementary skills. AI has collapsed the technical floor. What's actually scarce is user trust. The female partner's audience account is the optimal answer.

MCP One Year On: The Protocol Won, Vendor Moats Evaporated

Mon, 18 May 2026 00:00:00 GMT

A year ago MCP launched and everyone was betting on whether it would become another forgotten protocol. Verdict's in: it won. But winning meant flattening every IDE vendor's differentiation. Cursor can call Claude Code's servers, Windsurf can call Cursor's, tool wars degenerated from 'what integrations are exclusive to me' to 'whose skin looks nicer.'

AI Amplifies the Mid-Career, Not the Junior — A Counterintuitive Truth

Sun, 17 May 2026 00:00:00 GMT

The default narrative says AI takes out mid-career first. Pull the data apart and the direction reverses — AI favors juniors on narrow, scripted tasks; it overwhelmingly favors mid-career on open-ended, ambiguous ones. Judgment, failure-smell, system thinking, organizational fluency: these four experience assets aren't replaced in the AI era. They get multiplied.

Code Review Is the Real AI-Era Bottleneck — And Your Team Isn't Doing It

Invalid Date

GitClear 2024 data: 8x growth in duplicated code blocks, rising short-term churn, refactoring share down from 25% to <10%. DORA 2024: AI teams' delivery stability is down 7.2%. My own 12 EM 1:1s match: PR size grew 2-3x, review time actually went *up* (AI code is harder to read), but 'approved without comments' moved from occasional to default. Quality debt gets pushed 2-3 quarters out.

Digital Coders Explored: An Overhyped Narrative on Top of Under-Built Engineering

Fri, 15 May 2026 00:00:00 GMT

「Our AI digital engineer is live」 posts are everywhere. Look closer: 90% is a chatbot with an avatar. The engineering underneath — persistent identity, long-term memory, autonomous loop, collaboration interfaces — is the real lever for organizational productivity in 2026-2028.

Backend Falls First, Not Frontend: The Real Order of AI Replacement Across the Stack

Thu, 14 May 2026 00:00:00 GMT

Everyone says frontend gets killed first. Wrong — AI eats backend first, native client last. Replacement speed isn't about code difficulty; it's about whether the full work loop closes in text.

AI Coding's Real Bottleneck in 2026: From Tokens to Your Physical Workstation

Wed, 13 May 2026 00:00:00 GMT

Tokens stopped being the AI Coding bottleneck. In 2026 what slows you down is whether your monitors fit 5 active streams, whether your RAM holds 3 agents in parallel, and whether your network handles every API round-trip reliably.

Cursor's $20B Peak: Users Are Quietly Migrating, And You Should Know Where

Tue, 12 May 2026 00:00:00 GMT

March 2026: Cursor closed its $20B Series D. Same window: DAU retention fell from 76% to 68%. That's not a contradiction — it's 'valuation prices the last 6 months, users vote with the next 6.' Track where they're going, and you can read 2027's tool landscape today.

Measuring AI Coding Productivity: Why 5x Coding Doesn't Mean 5x Delivery

Mon, 11 May 2026 00:00:00 GMT

Mainstream AI Coding metrics only measure the coding segment, missing the full pipeline. Breaking the SDLC into 6 phases shows real productivity gain is 15-30% (not 5x), and the bottleneck has moved to Review and testing.

Vibe Coding's Security Debt: The Bill Comes Due in H2 2026

Sun, 10 May 2026 00:00:00 GMT

Vibe coding lets PMs, designers, ops 'just ship something.' But AI-generated code carries three systemic vuln classes by default — SQL injection frequency is *higher* than in 2020, hardcoded secrets appear in 31% of vibe-coded repos, OAuth misimplementation remains a classic. The debt is already on the books. Settlement comes in H2 2026.

Chinese LLMs Won on Cost: DeepSeek + Qwen + Kimi at 1/10 the Per-Task Price of GPT/Claude

Fri, 08 May 2026 00:00:00 GMT

DeepSeek-Coder-V3, Qwen3-Max-Coder, and Kimi K2 deliver AI Coding at 1/10 the per-task cost of GPT-5.5 / Claude 4.x — with SWE-bench Verified scores within 10%. In enterprise batch-workload scenarios, this gap is disruptive. Western vendors are responding slowly. This isn't nationalism — it's an engineering ledger.

Local LLMs Are Already Daily-Driver Ready: M4 + GGUF, No Cloud Needed

Thu, 07 May 2026 00:00:00 GMT

'Local models are always 2 generations behind frontier' — that was true in 2024. Not in 2026. Qwen3-Coder 14B at Q5, on an M4 Pro: daily AI coding tasks at parity with GPT-5.5, and 30% of scenarios are *faster* (no network round-trip). Privacy, offline, zero marginal cost. The hardware wall is gone.

Junior Onboarding Is Broken: AI Stole Mentor Time, Nobody's Coaching the New Hires

Wed, 06 May 2026 00:00:00 GMT

AI doubles mid-level productivity — at the cost of zero mentoring time. Across 9 companies' onboarding data, junior 'first-year growth curve' is collapsing: time-to-independent-contribution stretched from 4 to 7 months, 3-year voluntary attrition jumped 90%. This is organizational damage that compounds and won't fully surface until 2027.

$10M ARR with 3 People: The Real Structure of Hyper-Solo Companies

Tue, 05 May 2026 00:00:00 GMT

'3-person team, $10M ARR' was hype in 2023. By 2026, there are 20+ documented cases. But these aren't 'solo hero' stories — they share a precise structural recipe: 1 AI-leveraged engineer + 1 distribution lead + 1 customer success, with 8-12 AI agents replacing functional roles. See the structure or the narrative will fool you.

Leetcode Is Dead, But 90% of Companies Still Ask Leetcode

Mon, 04 May 2026 00:00:00 GMT

AI solves leetcode in 5 seconds. Yet in 2026, 90% of companies still ask vanilla leetcode questions. This isn't 'enterprises are slow' — it's interviewers pretending AI doesn't exist. The result is collective cognitive misalignment, filtering out the genuinely strong and selecting for test-takers.

The 'Ferrari Uncle' and 45 Years on One Craft: Senior Engineers Are Being Repriced in the AI Era

Mon, 04 May 2026 00:00:00 GMT

A 57-year-old dance professor goes viral for out-dancing a crowd of judges in suits, and everyone files it under 'midlife comeback.' It isn't. It's a market signal: when production cost collapses, judgment becomes the entire price. Shallow skills keep depreciating, deep judgment keeps appreciating, and the gap is widening — plus what actually happened when I rolled out AI Coding across a 20-person team.

The Token Economy Collapse: From Pay-Per-Use to Unlimited Subscriptions — Who Pays the Subsidy

Sun, 03 May 2026 00:00:00 GMT

In 2024 everyone counted tokens. By 2026, 'Claude Code Max $200/mo unlimited', 'Codex Pro all-you-can-eat', 'Cursor unlimited' are mainstream. The user wins — token-counting is over. Behind the scenes, four parties are losing: Anthropic, OpenAI, mid-tier tooling vendors, and heavy users themselves. See who subsidizes whom, and you'll see how this game ends.

Counteroffer vs. Jumping: What the Five-Year Math Actually Says

Sat, 02 May 2026 00:00:00 GMT

The 'I switched jobs and got a 30% raise' story travels furthest and misleads most. Judging an offer by how much the base moves is a rookie calculation — the real return only shows up when you run three to five years of total compensation. Counteroffers aren't automatically the trap the internet says they are, either.

Why You're Not Getting Promoted: The Three Months Before the Packet

Sat, 02 May 2026 00:00:00 GMT

90% of 'I can't get promoted' is not a capability problem — it's a narrative problem. A year of your work turns into five bullet points on a slide, and the committee is evaluating whether those five lines make them remember you. Doing the work and getting it told are two separate jobs; most people start the second one the week before the deadline.

AI Coding Flips English Disadvantage: How Non-Native Developers Are Quietly Winning

Sat, 02 May 2026 00:00:00 GMT

For 15 years, Chinese developers hit a ceiling on English: slow at reading docs, weak at commit messages, awkward in interviews. AI just demolished that wall. Claude writes your PR descriptions, parses papers, mediates 1:1 conversations — at a quality higher than 95% of native speakers. **This is a wildly underestimated reverse-tailwind window**.

Getting Out: The 18-Month Path from Decision to Exit

Sat, 02 May 2026 00:00:00 GMT

Somewhere around forty, tech veterans start thinking about getting out. The ones who leave with any dignity give themselves an 18-month runway from the moment they decide; the impulse version — fed up on Tuesday, resigned on Wednesday — mostly turns into regret inside six months. Leaving isn't a single move: it's money, skills, relationships, and your head, adjusted in parallel.

Reading the Next Industry Shift Two Years Early

Sat, 02 May 2026 00:00:00 GMT

The Chinese internet has flipped several times in thirty years: PC to mobile to O2O to short video to large models. Every flip broadcast its signals about two years ahead, and most people in the industry didn't notice until the coverage was everywhere. Two years early and you have time to move; two years late and you take whatever comes — the pay cut, the layoff, the career change you didn't choose.

Your First Month: How to Build Trust With a New Manager

Sat, 02 May 2026 00:00:00 GMT

A manager's read on a new hire is roughly 90% locked in within four weeks, and once they've decided 'this one needs chasing,' no amount of later diligence really overwrites it. The conversational rhythm you set in those first weeks defines how they use you for the next year — which is why treating month one as a waiting room is the most expensive new-hire mistake.

Six Months In and Already Want Out: Real Signal, or Just the Adjustment?

Sat, 02 May 2026 00:00:00 GMT

Of the people who want out six months into their first job, at least 70% are misreading 'I haven't adjusted' as 'I picked wrong.' New grads arrive expecting a curriculum, a manager who explains things patiently, and teammates who help each other out — none of the three hold anywhere. Three questions that separate a real signal from the adjustment.

Manager or IC: Three Questions to Ask Yourself Honestly

Sat, 02 May 2026 00:00:00 GMT

The IC track and the management track fork around senior/staff, and most people make the choice passively — the boss asks you to take the team, or nobody else is a plausible candidate. The costliest assumption is that management is the higher rung, a belief that runs especially deep in Chinese tech. They're parallel tracks: different work, different wiring.

Stuck Below Staff: The Jump from Executing to Setting Direction

Sat, 02 May 2026 00:00:00 GMT

Senior to staff is the hardest step on the IC ladder — around 70% of people sit here for two or three years, and some never get through and quietly stop trying. The reason usually isn't that they're not technical enough. It's that nobody told them the job changed identity: from executing well to deciding what gets executed.