Build a Disciplined Coding Agent

• ✓Build a plan-gate and a critique-gate for a coding agent, and wire them into one toggleable loop
• ✓Write a plan rubric that rejects confident-sounding filler and only passes a plan that names real edge cases
• ✓Build a critique step that reviews a draft without ever peeking at the hidden test oracle
• ✓Measure what discipline actually costs and buys — pass rate, first-pass rate, iterations, wasted edits — on a fixed task suite, gates on vs. off
• ✓Run the loop against a real local model and diagnose a failure mode a scripted model can never show you
• ✓Package the working loop as a reusable Claude Code skill

Build the habit behind every reliable coding agent — plan before you write, critique before you ship — and prove with real numbers what it costs and buys, not just claim it helps.