AGENT DESIGN · WORKFLOW ENFORCEMENT
Addy Osmani's agent-skills library, now past 79k stars, ships production SDLC skills that do not just tell a coding agent to write tests or review its own diff. Each skill carries a table of the excuses agents reach for to skip that step, with a rebuttal for each one, plus concrete evidence the agent has to produce before it is allowed to move on. This is a small working demo of that shape, no real LLM behind it, just the interaction.
Agent tries to shortcut a build
Four checkpoints from a real SDLC run. At each one, see the excuse an agent reaches for, then the rebuttal, then the evidence it has to produce before the checkpoint clears.
DEFINE
VERIFY
REVIEW
SHIP
WHY THIS MATTERS FOR CLIENT WORK
Left alone, a coding agent optimizes for the shortest path to a green checkmark, not for the invisible work senior engineers do by habit: writing the spec, actually running the change, rereading the diff cold, thinking through the rollback. Telling an agent to do those things once, in a prompt, gets rationalized away the first time it is inconvenient. What holds is a checkpoint with a named excuse, a rebuttal already written for it, and a real evidence requirement before the step is marked done. I run Claude Code on client builds daily. This is the difference between an agent that feels reliable and one that just sounds reliable.
Built 19 July 2026 · pattern sourced from Addy Osmani's agent-skills