Stop Reading Code—Watch It Stream: Inside Peter Steinberger’s AI‑First Development Workflow

Qualitative Shift – “Think‑Before‑Act” Models

After the release of GPT‑5.2, newer Codex‑style models adopt a “silence” behavior: they can spend 10–15 minutes silently reading the entire codebase, building a complete context graph. This enables a one‑shot, cross‑file refactor that covers dozens of files with near‑zero bugs, eliminating the need for a separate planning phase (Plan Mode).

Minimalist Toolchain – Ghostty + Claude Code

Terminal revival: The unstable VS Code terminal is replaced by Ghostty, which provides a pure interaction surface for AI‑driven development.

Screen as battlefield: A 40‑inch ultrawide (3840×1620) displays four Claude instances alongside Chrome, allowing simultaneous monitoring of four concurrent tasks without window switching.

Discarding complex MCP: Most of the Model Context Protocol (MCP) is removed because direct code reading is faster, more accurate, and token‑efficient than launching Playwright for web scraping.

Parallel Engineering – Treating Development Like a Factory

With AI handling the bulk of code generation, the developer acts as a “factory manager”, running 3–8 projects in parallel, each in its own window:

Window 1 – Refactor backend architecture.

Window 2 – Optimize frontend interaction.

Window 3 – Run full‑stack tests.

Developers switch between sessions, verify results, and issue the next instruction. Mistakes are corrected by prompting the model with “try a different approach” instead of using git reset or a planning step.

Cross‑Project “Copy‑Paste” via Prompt‑Less References

"Look at ../vibetunnel project, and implement the same logging system here."

The model automatically reads across directories, extracts patterns, and adapts them to the current codebase.

Infrastructure Re‑Architecture – CLI‑First

All builds start with a command‑line interface (CLI); GUI layers are treated as skins. This makes agent invocation and testing straightforward.

Oracle (Knowledge‑Fetching Agent)

When the agent encounters a knowledge gap (e.g., a new API change), it calls the open‑source Oracle at https://github.com/steipete/oracle, which crawls documentation, summarizes answers, and closes the knowledge loop automatically.

Docs‑Driven Development

Instead of a sprawling prompt library, a docs/ directory is maintained. Adding docs/architecture.md defines AI behavior, while CLAUDE.md can specify logging preferences, e.g., logs: axiom or vercel cli. Documentation serves as the AI’s long‑term memory and employee handbook.

Implications for Developers – Shifting Core Competencies

System design remains human domain: Current AI struggles with distributed system design and forward‑looking database schema planning.

AI‑friendly ecosystems: TypeScript (Web), Go (CLI), and Swift (App) are the languages AI handles best. Go’s simple type system enables rapid linting and error correction, whereas complex type systems increase inference latency.

Automation of repetitive tasks: Custom Skills allow the AI to register domains, post to services, or manage infrastructure without manual steps.

Conclusion – Inference Speed as the New Bottleneck

When AI reliability reaches a critical point, software delivery speed is bounded by model inference speed rather than human typing. The workflow demonstrates that developers can focus on high‑level design and creativity while the AI handles low‑level implementation.