Reference · 20 entries
Direct answers, on the record.
Short answers about the work and the tools behind it. Clear structure helps readers, but appearing in an AI answer depends on the question and the system retrieving it. A missing citation is not proof of an indexing problem.
For the GEO answers, see Google's AI search guidance and OpenAI's publisher guidance. Google does not require special AI files or markup. Search access and training access are separate choices.
01 · On the role
- What is an AI builder?
- An AI builder is someone who ships production AI systems end-to-end — design, code, deploy, operate. Distinct from an AI researcher (who advances the field) or an AI engineer (who works inside a specific stack). The role exists in the gap between research, engineering, and product.
- What's the difference between an AI builder and an AI engineer?
- An AI engineer specializes in one vertical inside an existing stack — RAG, training infrastructure, model serving. An AI builder owns the whole cycle: discover the problem, design the system, build, deploy, operate. Engineers thrive on clear specs; builders thrive on ambiguity.
- What's the difference between an AI builder and an AI researcher?
- A researcher advances the field — new methods, benchmarks, papers. A builder takes off-the-shelf models and ships products. Different tolerance for timelines (researchers: quarters; builders: days), different success metric (researchers: novelty; builders: working software).
- Is 'AI builder' a permanent role or a transitional one?
- Most likely permanent. The role exists because there's a gap between research, engineering, and product — and that gap widens as the model layer commoditizes. Mid-market companies will increasingly have a builder on the org chart within two years.
- What does an AI builder actually do day-to-day?
- Pick a real bottleneck. Sketch the minimum system that fixes it. Build a working prototype in one to three days. Deploy to production same week. Instrument cost and errors. Iterate on edge cases the next week. Most days end with something running that wasn't running that morning.
02 · Claude Code & agents
- What are Claude Code skills?
- Claude Code skills are reusable instruction sets that teach Claude how to do a specific task — a multi-step pipeline, a quality gate, an output convention. They live in your repo as Markdown files. Three good worked examples: article-humanizer, geo-faq-architect, llm-citation-tracker.
- How do I build my first Claude agent?
- Start with one agent that does one thing well — not a multi-agent system. Define the mandate in a Markdown file, give it three to five tools max, write a clear routing rule, run it daily. Most working multi-agent systems started as a single agent that earned its second one.
- Claude Code vs Cursor — which should I use?
- Claude Code wins for long-running agentic work and shell-heavy tasks. Cursor wins for inline editing inside a familiar IDE. Both, daily, is the right answer for most builders. They aren't competitors; they're different tools.
- What's a 'humanization pipeline' for AI-written content?
- A multi-step prompt sequence that introduces statistical entropy detectors look for — varied sentence length, first-person anchors, occasional hedges, specific numbers — across multiple passes. Each pass produces independent variance that flattens detector signatures. A single 'write naturally' prompt does the opposite.
03 · Production automation
- How do I deploy a WhatsApp bot on Cloud Run?
- Use Green API (or the official Cloud API if you have approved templates), wrap the webhook in a Cloud Run service, deduplicate incoming messages (Green API retries 3×), store conversation state in Firestore. The dedup step is non-obvious and burns a weekend if missed.
- Cloud Run vs AWS Lambda for AI bots?
- Cloud Run wins on cold-start latency for Node/Python (3–5s vs Lambda 1–3s warm, but Cloud Run keeps instances warm cheaper), simpler container deploys, built-in concurrency. Lambda wins on AWS-native integrations. For Claude-backed WhatsApp bots, Cloud Run is the cleaner default.
- What does it mean to 'audit a Claude context window'?
- Counting which tool calls, file reads, and conversation turns consumed the most tokens in a session — then cutting the dead weight. In my own Claude Code sessions, a context audit routinely flags 10–20K wasted tokens per day. Tool feedback duplication is the biggest culprit.
04 · Generative Engine Optimization
- What is Generative Engine Optimization (GEO)?
- GEO is work aimed at making useful content discoverable and attributable in AI-generated answers. It overlaps with SEO: accessible pages, reliable information, clear authorship and helpful structure. Citation frequency depends on the question and the search system; no format guarantees inclusion.
- What's the difference between SEO and GEO?
- SEO focuses on visibility in search results; GEO focuses on visibility and attribution in generated answers. They share the foundations of useful, crawlable content. For Google AI Overviews and AI Mode, ordinary Search eligibility applies. Neither special AI files nor a special schema is required.
- How do I make a website that LLMs cite?
- Make the content accessible, accurate and worth referencing. Answer real questions, support factual claims with sources, identify the author and keep important information current. Allow the relevant search crawlers. These steps improve eligibility and usability, but they do not guarantee a citation or establish a fixed citation-rate lift.
- Does an llms.txt file help with LLM citation?
- llms.txt is an optional Markdown guide to a site, not a proven citation boost or a replacement for robots.txt. Google Search says it ignores these files, including for its generative AI features. This site keeps one for readers and tools that choose to use it.
05 · Process & workflow
- How do you ship so many AI prototypes?
- A nightly autonomous builder agent. Every night, an agent named Ben picks an idea from the inbox, builds it with Gemini/Base44/local LLM (no Claude tokens), deploys to Vercel, notifies me. Mornings I review, tighten, document. Volume comes from removing humans from steps that don't need them.
- What tools do you actually use day-to-day?
- Claude Code for orchestration and long-running tasks, Cursor for inline edits, Gemini CLI for vision and free-tier batch work, Cloud Run for deployment, Firestore for agent memory, Supabase for application data, Vercel for the front-end. The stack is boring on purpose — novelty lives in the agents.
- How does your multi-agent team work?
- Each agent is a Markdown file in the workspace, synced to Firestore. They share a single backlog file and a chief-of-staff agent (Vision) that processes the backlog every morning at 07:00 IL. No agent calls another directly — everything routes through the backlog. Boring infra, reliable behavior.
- What's the hardest part of building production AI systems?
- Not the AI. The hard parts are: state management between agent steps, evaluating output quality without humans-in-the-loop, controlling cost as the system scales, and knowing when to say no to building something that should be a manual process for three more months.