Agent Skills
Addy Osmani introduces 'Agent Skills,' a framework to infuse AI coding agents with senior engineering discipline, guiding them through crucial steps like specs, testing, and reviews. This popular approach aims to prevent AI from acting like an overly eager junior engineer, ensuring reliable software delivery. HN discusses whether such elaborate scaffolding is truly necessary or if it adds complexity, and compares it to other agent frameworks like Superpowers.
The Lowdown
Addy Osmani's 'Agent Skills' project addresses a critical gap in AI coding agents: their tendency to bypass the diligent, often 'invisible' work that senior engineers perform to ensure robust, reliable software. These agents, like eager junior developers, prioritize task completion over the comprehensive processes of specification, testing, and thoughtful review.
- The Problem: AI agents naturally take the shortest path to 'done,' skipping essential senior engineering steps (specs, tests, reviews, scope management) that prevent incidents and ensure quality.
- The Solution: Agent Skills: A framework of 'skills' (markdown files with workflows, not just prose) injected into the agent's context to enforce these crucial engineering practices.
- Core Principles: Emphasizes 'process over prose,' utilizes 'anti-rationalization tables' to pre-empt agent excuses, mandates concrete 'verification' for task completion, employs 'progressive disclosure' to manage context size, and enforces 'scope discipline.'
- Google's Influence: The skills are deeply rooted in Google's engineering culture and practices (e.g., Hyrum's Law, test pyramid, 100-line PRs, Chesterton's Fence), applying well-known principles where agents typically default to skipping them.
- Implementation & Portability: Can be installed via marketplace (Claude Code), manually added to various tools, or simply used as a conceptual guide. The markdown format ensures portability across platforms.
- Broader Context: Skills are a component of 'agent harness engineering,' especially vital for long-running agents where shortcuts amplify problems.
'Agent Skills' proposes that the 'senior-engineer parts of the job are no longer optional, even when the engineer is a model.' By institutionalizing discipline and a structured workflow, the project aims to elevate AI coding agents from mere code generators to contributors of high-quality, maintainable software.
The Gossip
Scaffolding Scrutiny
Many commentators expressed skepticism regarding the necessity and effectiveness of elaborate agent scaffolding systems like 'Agent Skills.' Critics questioned if such detailed workflows truly improve outcomes beyond basic prompting, with some suggesting these systems might be 'anti-patterns' or 'rube-goldberg machines' that overcomplicate agent interactions and consume excessive context. There was a debate on whether simple outcome-based prompting is superior to prescriptive process definition.
Comparing Code Companions
A recurring discussion involved comparing 'Agent Skills' with other AI agent frameworks, particularly 'Superpowers.' Users inquired about their differences, adoption rates, and practical efficacy. While some praised 'Agent Skills' for its structured approach, others shared negative experiences with 'Superpowers,' citing increased token usage and slowed performance compared to vanilla agent setups.
Practical Praise and Pilfering Principles
Despite some reservations, many users found the concepts presented in 'Agent Skills' highly valuable. Several expressed intentions to 'steal' or adapt the underlying principles, such as anti-rationalization tables and the emphasis on verification, for their own engineering teams or customized agent setups. Users shared positive experiences, noting how the framework helped them focus on high-level architecture and product design, and improved the quality of agent-generated code.