What we broke, measured and fixed while building things: security, AI agents, WordPress and apps.
I rewrote a plugin listing into Portuguese, committed it, and the directory served the old English title for forty minutes while reporting it had been updated. A trunk commit moves “Last updated” and changes nothing anyone can see. Then I got the diagnosis wrong in the other direction.
A free plugin sat in the WordPress.org directory for 35 days and 195 downloads before anything — its code, its readme, or our own site — said a paid version existed. The readme fix needed no release and no review, and it still went last.
A readme-only commit traded the tag brazil (602 competing plugins, we ranked 150th) for validar cpf (9 competitors). Ten hours later we were 4th. Measured against 5,304 plugins of the same age, here is what that win is worth — and what it is not.
A guard that stopped the overnight suite at 85 °C was reading max() over the /sys thermal zones — a chassis probe up to 26 °C above the die. It killed 8 of 14 runs while the CPU sat in the fifties, and systemd recorded every one as a success.
A screen reader opens a PDF by reading its title and its language. Most set neither, so it announces the filename and guesses the pronunciation — and the harder thing those two one-line fixes are not.
A rough invoice sits unpaid because it is missing the fields a payer files it against. What turns a screenshot into a document AP can process — and the one thing reformatting cannot invent.
Dragging the browser thin renders the desktop site smaller, not the mobile layout. The three things a real phone capture has — viewport, pixel ratio, load time — and the one page a service cannot reach.
Seven packages ship inside one plugin zip under five licences; our composer.json names two. One command reads the real list, and “or later” in your plugin header is the trap rather than the escape.
The marker every tutorial tells you to grep for is a declaration, not an inventory: across 120 files on one disk, 30 claimed a form and 5 had one. The two commands that tell you whether any tool can make your PDF fillable.
One link added to the footer touched 82 pages, and 61 of 121 sitemap URLs began reporting that day as lastmod. Telling a sweep apart from a real edit needs one constant, and I measured it over 34 site-wide commits rather than guessing.
Our chess variant rests on one rule: a knight’s check cannot be blocked. Saying it on a page taught nobody, so the page now deals you the position instead — and the three positions I composed by hand were all rejected by the checker, one of them as illegal.
App Review takes a day or two and tells you nothing while it runs. A small script on the App Store Connect API fixes that — and four non-obvious things will quietly break it, every one of them silently.
Carlyle said history is the biography of great men; Spencer said the men were made by their times. We wrote down 26 game ideas in one afternoon and 18 already existed. AI makes ideas nearly free, so the old argument has a new answer: the credit is moving from the person with the idea to the filter that decides which ideas die.
We made escrow the default behaviour of a token instead of a contract you opt into: a transfer holds for seven days before the receiver can claim it. The escrow half shipped to Arbitrum One. The dispute half never came online, because routing a case to a Kleros court needs a whitelist that is governor-only — and we never asked before building. The design, what it breaks, and the three-bucket dependency check that would have caught it.
Your company’s name is on the subcontractor’s certificate, so you’re covered — right? The certificate holder box only means you got a copy. What actually puts you on their policy is a different field entirely, and how to tell them apart on the form.
Playwright’s networkidle waits for 500ms of total network silence — which a page that polls, beacons analytics or holds a socket open never reaches, so it times out on every navigation. Why a fixed sleep is worse in both directions, and the two signals that tell you a page is actually ready to drive.
A CPF and CNPJ carry their own check digits, so most mistyped Brazilian tax IDs can be caught in the browser before checkout. The mod-11 algorithm for both, the repeated-digit number that beats it, and what a valid ID still cannot prove.
Gate your best posts and their search traffic quietly collapses. Why serving Googlebot the full text is the trap that gets a site penalised, and the structured data that keeps gated pages findable.
An AI agent that types a password puts it in the transcript, where it is logged, cached and sent to the model on every turn, and reads it back the moment it checks the field. Why scrubbing logs is the wrong layer, and how to hand the agent a key name instead of the secret.
Your script clicks Save, the dialog closes, the code moves on — and the row was never written. Why a click that returns success is not proof the action happened, and the one check that turns a silent failure into an honest one.
Fetch.ai’s uAgents framework catches every handler exception and writes it to a log — so under pytest, a handler that crashes on every message still shows a green test. Why testing a uAgent is awkward, and how to test one offline: no Almanac, no wallet, no sleep().
Some logins fight scripts on purpose — 2FA, a captcha, a device check. Instead of reproducing the login, log in once by hand and attach Playwright to that Chrome over CDP with connectOverCDP. The one line that logs you straight back out, and the security catch nobody mentions.
WooCommerce makes you open every order to print its invoice — there's no “print all” button, even with an invoice plugin. Why that gap exists, and the four things a real bulk-print action has to get right (including the HPOS hook that breaks half of installs).
Playwright says the element isn't clickable, or the click silently hits a cookie banner and the test fails three steps later. Why it happens, why force: true makes it worse, and the hit-test fix that clears any overlay.
The Zora Coins API reports a market cap of zero for most coins, even ones with a live price — the field only fills when the pool trades against ZORA. Why it's blank, and the one-line derivation that gives you the real number.
A booking network sells discovery, then charges a fee for every diner it seats — including the regular who already knew where to find you. When that trade is worth it, when it isn't, and how to run reservations and no-show deposits in WordPress.
Playwright's MCP server hands a model 128,000 tokens to read five pages it has never seen — almost none of it the element it came to click. Where that cost comes from, the other ways an unseen page fights back, and how we cut the page 44× smaller.
No CPF, no número or bairro, and a CEP that autofills the street but leaves the store's state wrong. What a Brazilian buyer actually needs at checkout, and the one autofill bug behind most of it.
A JSON formatter, a JWT decoder, a Base64 site: the tabs you leave the editor for cost you your flow and, sometimes, a copy of your data on a stranger's server. The free ways to do each locally, and where an offline extension picks up the slack.
Past thirty applications you lose the thread: who to follow up with, which posting wanted what. The six fields that carry the whole hunt, and when a spreadsheet beats an app.
A one-line ceiling made six of seven levels uncrossable. The hard gate never saw it, because it computed jump arcs in open sky. What it costs when a test models the physics instead of reading it.
railway up Actually Uploads
Nine commits sat invisible to users for seven hours while every deploy reported success. Two facts about Railway deploys worth knowing before you automate one.
A background coding agent reintroduced a known game-breaking bug while every gate reported green. Why a stale base is worse than a merge conflict, and the one command that catches it.
Across 155 Zora creators, those whose biggest audience is on Farcaster have smaller followings and five times the holders of X-led creators. What the numbers say, what they don’t, and what a creator can do with them.
Zora’s reward event has no indexed fields, so a node can’t filter it by your address. Fetch every event, read the right words out of the data, and total your creator and referral earnings, with code that runs.
Apple rejects new developers with a six-item questionnaire and a demand for a recording from a real device. What each item is really asking, the empty in-app purchase list that turns your next submission into a 3.1.1 rejection, and what recording on real hardware actually costs.
A $500,000 buy costs 2.21% in AAPLx and 0.01% in SPYx. Seven days of Jupiter quotes across eight tokenized stocks, and the on-chain depth turns out to have nothing to do with the real stock.
What the organiser fee actually buys, why per-ticket costs hurt cheap community events most, and the one-event test to run first.
What the 15–30% actually buys, the arithmetic to run before you switch, and when the delivery platform is still the right answer.
QR tickets, phone-based door check-in, and the ordering mistake that quietly breaks homemade ticketing.
Four levels of access, why WordPress roles are the wrong tool, and what should happen when a membership lapses.
Codemagic or a git-hosted macOS runner, plus the OpenSSL trick for generating the signing request without Keychain Access.
What Applicant Tracking Systems really do, the two real reasons resumes fail, and the mechanical fixes that work.
File type, columns, fonts and headings: the layout choices that decide whether an ATS can read your resume at all.
Getting the right tax breakdown onto the PDF: VAT for the EU/UK, GST for Australia/India, sales tax for the US.
Why order numbers aren’t valid invoice numbers in many countries, and how gap-free sequential numbering works.
Why non-Latin and right-to-left fonts come out as empty boxes in invoice PDFs, and the real fix.
Send the invoice automatically with the order email: which email to use, no code, and the numbering gotchas.
The wp.org-compliant free/Pro architecture, a reusable license gate, and how to actually sell the Pro version.
Add a paid Pro tier without breaking WordPress.org's rules: where the license check lives, how to verify a key, and how to fail open on a grace period so a customer's site never goes dark.
The free ways to test a regex, the edge cases that actually break patterns, and catastrophic backtracking, the failure mode that turns a one-line pattern into an outage.
The long way round: Kotlin, then C#, then React Native and Expo, and how to build, sign, and ship an iOS app to the App Store without owning a single Apple device.
Read a token’s header and payload, what the claims mean, why decoding isn’t verifying, and the security reason not to paste a live token into a random decoder.
A thirty-year veteran watches AI execute in four seconds what took a career to master, and feels obsolete. Why that machine metaphor is seductive, and exactly where it breaks.
The digital-nomad life looks like beaches. Behind the laptop it’s solitary survival: your funnel is a fish basket, marketing is casting into empty water, and the helicopter is always on standby.
A loop repeats a prompt. Recursion feeds a prompt its own output: refine, decompose, converge. What it is, why it beats a loop, and how to do it in Claude.
The internet runs on open-source maintained by a handful of unpaid volunteers, and AI just made the bill come due. The free-rider problem at the root of modern software.
AI made content infinite and nearly free, which flips what’s valuable from generation to proof: that a human made it, that a source is real, that a payment happened.
We gave AI agents the ability to think, plan, and code, then forgot to give them a way to pay. Why autonomous agents need their own wallets, and what breaks when they don’t.
The old rule was never ship code you can’t explain line by line. AI broke it. Here’s when shipping code you don’t fully understand is fine, and the exact line where it isn’t.
“Why hire a junior when AI writes the code?” is the most expensive short-term thinking in tech. Where senior engineers actually come from, and what happens to teams that stop growing their own.
Yes, objectively. A mostly-tongue-in-cheek defense of dark mode, and what the fight over it reveals about actually building software for humans.
Your photos carry hidden GPS coordinates and camera data. Why it’s a privacy risk, the free ways to strip it, and a $1 one-click cleaner.
Turn a stack of page photos into one real PDF: the free routes, their limits, and a $1 clean multi-page builder.
Why videos are huge, what actually shrinks them (bitrate, resolution, codec), and a $1 tool that hits a target size.
How upload validators check files, a 30-second pre-flight checklist, and a $1 tool that tells you before you upload.
The free phone and desktop ways to cut out a background, when they fall short, and a $1 one-click transparent PNG.
Why applicant tracking systems reject resumes before a human looks, the formatting and keyword fixes that get you through, and a $1 ATS check.
The real size, dimension, background, and format rules for ID photos, the free fixes, and a $1 tool that hits the exact spec.
The free script, the gotchas that break naive converters, and a clean $1 tool for CSV↔JSON in both directions.
Why HEIC uploads fail, what JPG actually fixes, and the fastest $1 conversion: no subscription, no account maze.
The upload box says 5 MB. Your file says 17. How to compress under the limit, and a $1 tool that does the loop for you.
Useless error messages, three usual culprits, and one $1 fixer that targets DMV, passport, and generic upload rules.
Honeypots, sell taxes, thin liquidity, whale wallets: what to check on Base, and a $1 risk scorecard.
Trailing commas, smart quotes, truncated objects: repair paths for humans and agents, including a $1 API.
No free trial maze. Card checkout for people, x402 for agents, same endpoints. The logic behind the $1 tools catalog.
The factors that move the price, realistic ranges, the ongoing costs nobody mentions, and how to spend less without building junk.
The documents to collect, the insurance and compliance to verify, and how to keep onboarding from bottlenecking your project start, before a sub sets foot on site.
A developer’s guide to charging AI agents per call: the 402 handshake, what a paid endpoint returns, and the fastest path to a machine-payable API without building billing.
A certificate can look perfect and still leave a general contractor exposed. The two endorsements that decide who really pays when a subcontractor causes a loss, in plain English.
The code that sat unused in the web’s rulebook for 30 years is now how software pays software. How 402 and x402 let AI agents pay for any API, no signup, no keys.
Tracking subcontractor Certificates of Insurance by hand quietly exposes general contractors to six-figure liability. What a COI protects, why manual tracking fails, and what to automate.
AI is marketed as multilingual, but its data, tokenizers, and concepts all lean English. How that bias works across three layers, and what it costs everyone else.
Your AI can reason brilliantly and read nothing on the live web. Here’s the one-line MCP fix that adds pay-per-call scraping and market data, no account, no subscription.
Why your scraper gets blocked, what residential-proxy unlockers actually do to get through, and how to pay for anti-bot fetches per success instead of per month.
A dormant HTTP error code is being revived so software can pay software. What the agentic economy means for you: the upside, and the real risks.
Cloud, Firebase, Play Console, AdMob, AdSense, Analytics: one app launch, seven dashboards, and config files holding it all together.
From Android Studio to Google Play: Kotlin, Jetpack Compose, Firebase Auth, Firestore, Cloud Storage, FCM, and CameraX, with the commands to build and ship your own.
Next.js, Prisma, Turso, NextAuth, Stripe subscriptions, Claude AI integration, and Railway deployment, step by step with every command.
IoT network security with Vigilo (75% IoT-23 detection @ 1% FPR) and predictive maintenance on C-MAPSS (100% on single-condition subsets); ~1.3M params, CPU-only.