Field notes from
the automation studio.

AI agents are starting to do the buying. We audited the structured data on our own two Shopify stores to see what an agent can actually read. Here's the gap.

Models got better at spotting crude injections, so people assume it's solved. It isn't. The risk moved from tricking a chatbot to hijacking an agent with tools.

67% use AI at work through personal accounts. 18% of companies have a policy. That gap is your biggest data liability, and August 2 removes the excuse.

Vercel fixed its Claude Code plugin after a privacy backlash. Our install never updated and kept sending shell commands for months. The lesson in stale installs.

GPT-5.6 (Sol, Terra, Luna) launched today. For most small businesses the version number is noise. Here's what actually changes, and what to ignore.

Sol, Terra, or Luna? Most people will default to the flagship and overpay. Here's which GPT-5.6 model fits which job, with real pricing and a benchmark surprise.

GPT-5.6 Sol makes subagents built-in with Ultra Mode. We've run agent fleets for years. Here's what it changes for business automation, and the trap to avoid.

Brave runs its own index, and getting into it is now an AI-visibility play. Here's how Brave Search actually discovers pages, and how to get yours in.

Agencies want $1–5K a month or 10–20% of spend. Here's what you should actually get for it — and why the creative, not the manager, decides your ROI.

Four AI meeting notetakers compared on what actually matters — bot visibility, CRM automation, and the two-tool combination power users actually run.

Perplexity, ChatGPT Search, and Google AI Mode compared on accuracy, citation patterns, and market share — and what each means for your AEO strategy.

Sora's API shuts down September 2026. Veo 3.1, Kling 3.0, and Seedance 2.0 compared on what actually matters now — and which one replaces your Sora workflow.

AI builders made DIY genuinely good. So when is doing it yourself smart — and when does it quietly cost more than hiring? An honest breakdown.

Claude Code, Codex CLI, and Gemini CLI ranked on real Terminal-Bench data, pricing, and 6 months of daily operator use across 4 businesses.

Three agentic browsers tested against real multi-step tasks, not scripted demos. Comet vs Atlas vs Genspark — which one actually books the flight.

Retell, Vapi, and Bland compared on latency, real pricing at 10,000 calls, and compliance. The voice AI stack we'd actually build a client's receptionist on.

We rolled out AI receptionists across 6 physio practices. No-show rates, after-hours capture, and what broke — the honest case study, not the pitch.

Anthropic's Fable 5 is globally available again after a 19-day export ban. Mythos 5 isn't. What the asymmetric restoration reveals about the new AI kill switch.

Half of small-business invoices go late. The automation that gets you paid faster costs less than a single overdue invoice — here's the setup.

Three dental AI receptionists compared on price, PMS integration, and real no-show numbers. Which one actually earns its cost, by practice profile.

The EU AI Act's 2026 deadline got pushed to 2027 and 2028. What small businesses actually still have to do now — in plain English, not law-firm language.

Not the fancy stuff. A priority order that starts with the automations that pay back in days — and the ones worth waiting months for.

WhatsApp banned general-purpose AI chatbots as of Jan 15, 2026. What's actually still allowed, what got shut off, and the rebuild path for small businesses.

Reddit for idea validation, Claude Code + Xcode 26 for the build. The 2026 iOS app pipeline — plus the .pbxproj rule that saves you a week.

Both recover the calls you miss. One texts back, one actually answers. Here's which recovers more revenue — and when the cheap option is enough.

The Reddit SEO courses sell three tricks. None are the real work. Five methods for Reddit-driven keyword research that actually earn rankings.

Reddit ads run 42% cheaper per click than Meta and 50–70% cheaper than Google Search. The 2026 small-business playbook, the CPM math, the traps.

Vendors swear customers love it. The data is sharper: people hate AI-only, not AI. Here's who actually asks for a human — and how to keep them.

What Reddit actually thinks about iPhone 18, iOS 27's Claude and Gemini extensions, and the 2027 anniversary iPhone. 200 threads, sorted by conviction.

The Fable 5 ban was sold as security policy. Cybersecurity experts argue it weakens defense without weakening attackers. The technical case against it.

The US can pull the strongest commercial AI from European hands in 4 hours. The geostrategic case for EU sovereign AI, after the Fable 5 export ban.

On June 12, the US Commerce Department killed Anthropic's two most powerful models in 96 hours. The timeline, the jailbreak that triggered it, what comes next.

Marketing budgets are trading positions. The automated CPA-threshold rules that kill losing campaigns in 60 minutes, not 60 hours — and the 30% data caveat.

The SaaS stack tax is eating your agency margin. Three internal micro-tools we built in a weekend each, the VPS economics, and when SaaS still wins.

Automate the 72-hour gap between Stripe payment and project kickoff. The n8n pipeline that provisions portal, Slack channel, and welcome sequence on its own.

Every tool a 3-person team actually needs in 2026, with line-item costs. The €500/month AI stack we run — and the tools we deliberately don't buy.

The 2026 answer engine optimization playbook. Six AEO levers, measured citation rates across ChatGPT and Perplexity, and the work that compounds.

AI phone receptionist for small business: rent at $25/month or build for $40–80. The honest math, the SaaS shortlist, and where each path lives or dies.

Seedance 2 + Hermes + Claude Code + OpenClaw — the agentic stack that runs a marketing campaign end-to-end. The pipeline, the math, the limits.

Claude Code now edits videos end-to-end — cuts, captions, motion graphics, sound. A 20-minute YouTube pipeline. What it means for the people who used to cut.

Workflow automation as a revenue driver, not a cost cutter. The taxonomy, the operator math, and the customer-facing AI that actually earns.

An SEO-ranked website for small local businesses needs 4 things, not 20. Here's the lean playbook — plus the LLM citation angle competitors miss.

Most of the 'one weird trick' TikTok ad hook tactics now flag as AI and tank reach. Here's what actually carries through 2026 — honest about what stopped working.

An operator-level breakdown of an AI ad creative pipeline shipping ~50 paid variants per week. The brief → image/video → upload → feedback loop, with real numbers.

Meta ads 2026, TikTok ads vs Facebook, Google search, Microsoft Ads — a platform-by-platform decision framework from an operator running budget across all four.

Two open-source multi-agent systems, two different bets. MiroFish simulates societies; OpenClaw executes work. Here's what each is for, and why it matters.

A field essay on what 'enterprise AI coding adoption' actually looks like in 2026 — past the headlines on Goldman Sachs, Stripe, Klarna, Shopify and the Microsoft Copilot Enterprise rollout. Who benefits, who gets displaced, what '10x faster' actually means, and what the next two years hold for the indie builders who watched this happen.

SEO and LLM citation are different games that happen on the same page. Here's the llm seo blog pattern that wins both in 2026 — structure, voice, proof.

A field report on the actual architecture behind a billion-view-per-month AI content pipeline — topic generation, Nano Banana Pro and Flux for stills, Kling 3 for image-to-video, Remotion and CapCut for assembly, LLM-as-judge for slop rejection, and the distribution layer that doesn't get accounts banned.

German-speaking markets are slower, pickier, and more profitable once you're in. Here's the dach ai agency pattern that actually wins — pricing, positioning, sales.

Most ai automation dach agency promises collapse on contact with a Mittelstand buyer. Here's the pattern that actually ships — scope, pricing, delivery.

OpenClaw shipped a reddit monitoring tool with Claude as the engineering backend in a single sprint. The full ai built saas case study — scope, security audit, GTM, and the first numbers.

I used all three every day for half a year across four businesses. Here's the claude code vs cursor verdict, the v0 vs cursor verdict, and what I wish I'd known.

A field report from running OpenClaw Factory — a 33-agent Claude architecture in production for 6 months. Multi-agent system production lessons.

Three months shipping kling 3 ad creatives for a DTC brand. What worked, what was expensive, what made Meta's algorithm happy — and what it didn't.

Twelve months of Meta ads for a kids-accessories Shopify brand. The meta ads kids products playbook — creative, structure, ROAS, and what doesn't work.

I watched two client sites lose 80% of organic traffic overnight. AI content SEO 2026 is not what the 'undetectable' tools promise — here's what Google actually penalizes.

Six months using nano banana pro brand imagery across four brands. The model's quirks, the prompts that work, the QA loop that keeps visuals on-brand.

Twelve months of a kids-accessories Shopify store, with every line item. The full wondrakids costs breakdown — ads, fulfillment, returns, software, founder time.

We pulled GemPages off a Shopify store and rebuilt every landing page as Horizon sections. Theme customization, Liquid custom sections, and the numbers.

Instantly runs the sends. v0 builds the prototypes. Claude handles the research. The instantly ai v0 stack in full — inbox warming, deliverability, the lot.

Sending a working prototype in cold emails raised our reply rate by 5x. The v0 cold outreach playbook — how we scope, build, send, and follow up.

How we use Claude Code as the operating system for four businesses at once — the claude code workflow, the hooks, and the AI agency solo founder tax.

The n8n vs claude agents question gets argued in ideology and decided in practice. Here's when a workflow beats an agent, and when it's the other way around.

I got tired of AI tools ignoring my design system. So I wrote a single file — taste-skill.md — that teaches v0 and Claude my ai design system prompt rules.

How we shipped a production MDX blog on Next.js 16 with Velite: the content layer, typed frontmatter, and why Velite vs Contentlayer was a short debate.

Every six months LinkedIn announces the death of the agency. I've run one for five years — the future of digital agencies, honestly.

From idea to paying customers in eight weeks. The ai trading education saas build log — stack, decisions, mistakes, and the 300-user beta that paid for itself.

Three years on Webflow. Shut down in a long weekend. The Webflow alternatives we shortlisted, the math behind the move, and the line item we underestimated.