Two client sites, three months, one rebranded AI builder. Where Blueprint AI Design Intelligence actually helps a designer — and where it defaults to the same template every other Squarespace site ships with.
Two real builds, one head-to-head: the same B2B SaaS lead-gen site, built first on Lovable (1 day, 6 credits, gorgeous) and then on WordPress (5 days, ranked page 2 in 30 days). What I'd pick for what, and why the SEO content flywheel is WordPress's real moat.
A 3-sub-agent competitive intel pipeline that produces an 8-page PDF brief every Friday at 6am — no human reads a single competitor page in between. The parent Claude agent dispatches a site-watcher, an ad-watcher, and a social-watcher, each returns a strict JSON schema, the parent synthesizes everything into a Markdown brief that Pandoc renders to PDF. The parent prompt, the three JSON contracts, the PDF template, and the failure modes that have actually cost me a brief.
Generate 100 paid-social ad variants per hour on a local M-series Mac via Ollama + Llama 3.3 70B, then run a second pass that scores each variant 0-100 on hook, value clarity, and CTA specificity. The full prompt pair, the Python wrapper, and the math that makes $0 marginal cost win over GPT-4o at any team size.
A production pipeline that takes a 'demo request' email, qualifies it against an ICP rubric, enriches from LinkedIn, books a Cal.com slot in the prospect's timezone, and sends a personalized prep doc 24h before the call — all before a human touches the lead. The full Make scenario, the actual Operator instruction set, the ICP rubric, the prep doc prompt, and the three failure modes that have actually cost me time. ⚠️ Operator UI may shift; the framework stays.
A 6-week, 8,400-email field test of running a privacy-respecting local-LLM email triage layer on Apple Silicon — Apple Mail → AppleScript → Ollama-served Mistral 7B or Llama 3.1 8B — with a 4-bucket rubric, real hardware benchmarks, and a 92% accuracy figure you can reproduce.
A 30-day blind test: 800 meta titles + 800 product descriptions rewritten by both Qwen 2.5 14B (self-hosted on a refurbished workstation) and Claude Sonnet 4, rated by 3 SEO contractors. The result is not a clean win for open weights — it's a split. Where Qwen breaks even, where it collapses, and the actual cost math behind self-hosting for repeat-pattern SEO work.
A production 3-model content pipeline where each AI does only what it is actually best at: ChatGPT drafts, Claude reviews, GPT-Image generates visuals. The actual hand-off prompts, the JSON contract between stages, the $0.41 / 18-minute cost and timing comparison vs single-model, and the model-identity-confusion failure mode that cost me a published post.
An n8n + Claude Haiku 4.5 agent that watches 9 subreddits, scores every post on a 3-axis rubric (intent, urgency, fit), and posts a top-5 daily digest to Slack — with the verbatim Claude prompt, the 9-sub shortlist, the cost math, and the week-2 finding that pruned it down to 4 subs for a +120% signal lift.
Self-hosting a 70B model sounds reckless for a marketing team. For 90% of teams it is. But there are 4 specific jobs — bulk ticket classification, private competitive intel, overnight SEO meta-generation, PII-redacted list cleaning — where the math flips and a single A100 + Ollama pays for itself in 4-7 months. Hardware reality, Docker compose, real throughput, and the 4 prompts.
A production n8n workflow that pulls decaying posts from GSC (Google Search Console, 谷歌站长工具), drafts a refresh brief, has ChatGPT write the update, validates the new draft, and ships it to the CMS — all in one loop. The full node map, the three prompts, the four validation gates, and the three failure modes that have actually cost me a post.
A working pipeline for dubbing paid-media video ads into 29 languages with ElevenLabs' Dubbing API — voice cloning, translation QA, and the lip-sync trap I learned the hard way.
