Build code with agents while keeping a working understanding
A developer describes a workflow where the agent interviews them on ambiguities, makes the minor decisions, generates explainers and quizzes, and runs verifiers in the coding loop.
Done with a custom-built or unnamed agent
- What they did
- Before building, the agent interviews the developer on ambiguities in the spec (a /grill-me style prompt). The developer makes decisions on stack and architecture only for differentiating parts and leaves minor details and commodity subsystems like login to the agent. A skill generates interactive explainers with quizzes. Linters, type checkers, pre-commit hooks, tests, formal-method traces and a headless-browser design verifier keep the agent checking its own work.
- How it went
- The developer says this keeps them out of the 2am fix-and-break loop while still building understanding. No measurements are given, and they note other engineers are split on whether understanding the code matters.
- Worth knowing
- The design verifier turns spec requirements into mechanical rules checked against DOM measurements in a headless browser, which the author says was impractical before agents.
Try it yourself with Claude Code
I want to build [feature or project] in this codebase, and I want to stay able to understand everything you write. Before coding, interview me about any ambiguities in the plan, one question at a time, and make the minor decisions yourself, listing them for me. After each major piece, write a short explainer of how it works and give me a 3-question quiz on it, then run the tests or checks before moving on.
Want help getting this working? Join the Vibe Coding Academy ↗ Run by Alex Finn, who also runs this site.
Source: HN · Oct 19, 2026
Related use cases
Build a C compiler with a team of parallel agents
Anthropic's engineering write-up on tasking Opus 4.6 agent teams with building a C compiler.
Build a mental models framework with a multi-agent dev team
A solo founder used Claude, ChatGPT, Windsurf and Cursor as role-based team members to design, test and deploy a 120-model mental-models framework and MCP server.
Migrate a SaaS from AWS to DigitalOcean
A developer used Claude Code as a pair programmer for provisioning, database seeding, SSL, DNS cutover and production debugging, finishing in about four hours with zero downtime by his account.
Run Terraform plan and apply through an MCP agent
The author had Cursor run Terraform against Postgres via MCP; the apply hit a permissions error and he finished from the terminal.
Crawl documentation into clean Markdown for AI agents
The author built a documentation crawler for their AI agents that pulled 60 pages of clean Markdown in 49 seconds.
Custom agent
Build an invitation card website purely by prompting
A user built a frontend-only invitation site with animations, images and navigation for his girlfriend, giving instructions to Cursor without touching the code.
Get 5 like this every morning
The best things people got an AI agent to do, each with the prompt to try it.