Agent Skills: bolting senior-engineer scaffolding onto AI coding agents

Addy Osmani argues that AI coding agents default to the same failure mode as junior engineers: take the shortest path to ‘task complete’ and skip the invisible work — specs, tests, scoped diffs, review-readiness, evidence of correctness — that actually separates reliable shipping from incidents. His Agent Skills project (now at 26K stars) tries to force that scaffolding back in by turning senior-engineer habits into agent-executable workflows organised around a six-phase SDLC: define, plan, build, verify, review, ship.

A ‘skill’ is deliberately narrow — a markdown file with frontmatter, injected into context when relevant, encoding a sequence of steps with checkpoints and exit criteria. Crucially, it is not reference prose. Essays on testing get read and ignored; a workflow that says write the failing test, run it, watch it fail, then implement gives the agent something to do and the human something to verify. A meta-skill router activates only the skills relevant to the current task, applying progressive disclosure so a twenty-skill library doesn’t poison a small context budget.

The most distinctive design choice is anti-rationalization tables: each skill ships with pre-written rebuttals to the plausible excuses an agent (or tired engineer) uses to skip steps — ‘too simple for a spec’, ‘I’ll write tests later’, ‘tests pass, ship it’. LLMs are fluent rationalisers, so the rebuttals are written before the lie. Combined with non-negotiable verification (every workflow ends in concrete evidence) and strict scope discipline (touch only what you were asked to touch, no drive-by refactors), the system encodes the same review norms Google and similar orgs already enforce on humans.