luxueliu-agent-discipline-skills

AI agents are never short on intelligence — they are short on discipline. Claiming completion when the file was never written, drifting off-task while still feeling confident, retrying the same failed hypothesis three times without converging, or overwriting a teammate's fresh write with a five-minute-old state in a multi-agent setup: none of this is fixed by prompt tricks. What is missing is an engineering-grade set of brakes and handoff protocols. luxueliu-agent-discipline-skills packages exactly that into four zero-dependency, single-file SKILL.md skills that work out of the box. deepcode-review enforces a disk read-back, a change inventory, and an encoding / escape scan before any completion claim; agent-course-correction halts momentum, restates the latest goal, classifies the drift type, and applies the minimal safe correction; systematic-debugging follows an evidence-driven loop — smallest failing sample, single hypothesis, regression proof — and stops at the root cause when only diagnosis is authorized; deepcode-collab guards multi-agent boundaries by re-reading before writes to prevent stale state, tagging gaps when inheriting half-finished context, and strictly separating executed from achieved. The four skills have explicitly stated, non-overlapping responsibilities, and every rule traces back to a real incident. This is for developers running multi-agent workflows in DeepSeek Harness, Claude Code, or Codex, and equally useful for human teams who want to bring the same discipline into code review and handoff processes — after all, the mistakes AI makes are exactly the mistakes humans make.