// AI_LAB · LIVE

How I Build With AI

Most of my work ships through a tight loop with Claude. The differentiator isn't the tool — it's the plan I bring to it. Below: the workflow, the tools, a few real case studies, and a live playground so you can try it yourself.

// LIVE_PLAYGROUND

Ask Claude about my work

This playground is briefed on me — my stack, my workflow, the kinds of projects I've built. Ask anything. Streams from Claude Sonnet 4.6. Eight prompts per visitor per day.

claude-sonnet-4-6 · live_playground.tsx

online

$ prompt

⌘↵ to run · 0/1200

$ claude.response

Awaiting your prompt. Pick an example above or type your own.

// THE_LOOP

How a brain fart becomes shipped code

Five steps, on a continuous loop. The diagram below cycles through each in turn — the same way the work does in practice.

Brain dump

/work-plan

Implement

Verify

Review

Brain dump

Start vague, on purpose

A new task almost never arrives as a clean spec. The first move is to get the half-formed idea out of my head and into Obsidian — bullets, a screenshot, a sentence. Structure comes later.

/work-plan

Turn the dump into a contract

My custom Claude Code skill takes the brain dump and produces a real plan: files to add, files to change, ordered steps, an explicit out-of-scope list, and a one-line definition of done. The plan is the contract — and the leverage.

Implement

Execute against the plan

With a real plan, the writing is the easy part. Claude handles the bulk, I redirect when something drifts, and if we deviate we update the plan before the code. The plan is the source of truth, not the diff.

Verify

Test it in the browser

Type checks prove code compiles. The browser proves the feature works. I click through the golden path, then sweep adjacent flows for regressions. Anything still broken loops back to step 03.

Review

Adversarial pass with BugBot

Cursor BugBot reads the diff with fresh eyes — catches what Claude and I both rationalized past. Anything flagged either gets fixed, or gets a deliberate decision logged before the commit lands.

the takeaway

Anyone can prompt Claude. Fewer structure the ask so well that the answer is almost predetermined.

— five steps, one loop, on repeat.

// THE_STACK

The tools, honestly

A short stack on purpose. I'd rather know a few tools deeply than dabble in twenty.

Claude Code

Primary IDE-grade agent

The workhorse. Lives in the terminal, sees the whole repo, runs my work-plan skill, fans out subagents. Almost every commit in this repo passed through it.

Cursor

Editor + inline AI

Where I write, read diffs, and do quick refactors. Tight loop between selection and prompt — the thing Claude Code can't do as well as a real editor can.

Cursor BugBot

Adversarial code review

Reviews every diff before it ships. Catches the silly stuff that Claude and I both rationalized past — unused effects, missing null guards, naive perf.

Obsidian

Thinking, before code

Where the vague idea becomes a working brain dump. By the time something reaches Claude, the rough shape already exists in a note.

// SUPPORTING_STACK

Claude APIproduction LLM calls

WarpAI-native terminal

T3 Chatmulti-model fast lookups

Supabasedatabase + auth + storage

Cloudflareedge + DNS + tunnels

// CASE_STUDIES

How I'd build it for you

Three AI-native patterns I work with most often. For each: the brief, the work-plan I'd open it with, and what would ship.

pattern · rag + vector-search

Semantic Search Over a Knowledge Base

THE_ASK

A growing team has Notion pages, Slack archives, and a wiki nobody can find anything in. Keyword search is useless. Need search that understands intent.

THE_PROMPT

/work-plan — design a RAG-style semantic search over a multi-source knowledge base. Embed Notion + Slack + wiki content into a vector store. Wrap a clean search UI on top. Constraint: low maintenance, runs on existing Postgres.

THE_RESULT

pgvector on the existing Postgres, Claude for both embeddings and rerank, single-bar Next.js search UI. Plan ruled out a separate vector DB before any infra was provisioned. Engineers find answers in seconds.

Claude APIpgvectorNext.jswork-plan skill

pattern · agent-fan-out

Multi-Agent PR Reviewer

THE_ASK

Pull requests get rubber-stamped under load. Need a tireless first-pass reviewer that catches bugs, suggests refactors, flags security issues — but doesn't block on style nits.

THE_PROMPT

/work-plan — automated PR reviewer using parallel Claude agents. One for correctness, one for security, one for refactor opportunities. Synthesize into a single PR comment. Skip style. Run as a GitHub Action.

THE_RESULT

GitHub Action fans out three Claude agents per PR, synthesizes a single structured comment in ~90 seconds. Catches real bugs the team missed in a meaningful share of pilot-week PRs.

Claude APIGitHub Actionsagent orchestrationwork-plan skill

pattern · embedded-claude

SaaS Onboarding Copilot

THE_ASK

New users churn before they hit the third screen of a complex product. Need a Claude-powered guide that explains features in plain language, adapts to the user's role.

THE_PROMPT

/work-plan — embed a Claude assistant in the onboarding flow. Briefed on product docs + user's selected role + their progress. Conversational, not modal. Streams responses. Lives in a sidebar.

THE_RESULT

Stateful sidebar chat with role-aware context and streaming responses on Vercel edge. Activation improved, first-week support tickets fell. Plan kept conversation history out of v1 — shipped in days, not weeks.

Claude APIstreamingNext.js Edgework-plan skill

Want this kind of AI fluency on your team?

Hire me to architect agents, embed Claude into your product, or train your team on a workflow that actually ships.

Start a conversation