How to do hard real-world coding with AI
A practical framework for using AI to complete non-trivial, multi-hour engineering tasks: move beyond toy prompts to a four-phase pipeline (Understanding → Planning → Implementing → Checking) with rigorous verification and iteration.
Sep 11, 2025 · Baseline, Dublin
The Augmented Software Engineer — Agentic-Coding vs. Vibe-Coding
Presentation
Talk overview
- Success criteria on two axes: accuracy (valid, idiomatic code; no regressions) and completeness (finish the end-to-end task).
- Four-phase pipeline: Understanding, Planning, Implementing, Checking — loop until checks pass.
- Prep the workspace with opinionated config files (AGENTS.md, CLAUDE.md, .cursorrules) and CI/type-checking.
- Use a project bootstrap script (./setup-project <name>) to create docs like REQUIREMENTS, DESIGN, VALIDATION_STRATEGY, PROGRESS_TRACKER.
- Verification: LLM-to-LLM self-consistency checks, strict TS type-checking, static analysis, tests, and logs everywhere.
- Get better outputs with model choice, prompt sampling + selection, orchestration into sub-tasks, tool use (e.g., calculator), and context engineering.
- Practical ops tips: persistent shells (e.g., screen) to run long flows; respawn agents automatically when checks fail.