Last week I shipped eight production websites — nealkhis.com, acs.nealkhis.com, gamblegrip.com, medcontact.ai, lab.nealkhis.com, hr.nealkhis.com, briefs.nealkhis.com, partners.nealkhis.com — plus a partner program backend, HR portal, investor relations page, and blog system. Seven days. One developer (me). No agency.
The core idea is what I call the Executor Pattern. Each major feature or site becomes a numbered executor script: run_e01.sh through run_e47.sh by the end of the week. Each executor contains a single Claude prompt with a precise task. The prompt is written in the script file, versioned in git, repeatable.
What makes it different from just prompting Claude in a chat: the prompts are structured as specification documents, not questions. They describe the exact files to read, the exact changes to make, the exact verification steps to run. Claude is not asked to figure out what you want. It is given a complete spec and executes it.
Parallel execution was the multiplier. Claude Code can run multiple tools simultaneously. A single executor that naively does file reads and writes sequentially might take 3 minutes. The same executor with parallel tool calls takes 40 seconds.
Total estimated cost: 40-50 hours of my focused time, plus API costs (roughly $15-20 for all Claude calls). An agency would have charged $40-80K for this scope and delivered in 8-12 weeks.
The part I underestimated: writing good executors is a skill. The first five I wrote were sloppy. By executor 30, they were precise enough that Claude almost never deviated from the spec. That learning curve is real but short.