## Sources

1. [Get a Good Return on Your AI Investments](https://www.oreilly.com/radar/get-a-good-return-on-your-ai-investments/)
2. [Agent Skills](https://www.oreilly.com/radar/agent-skills/)

---

### **Agent Skills**
**Author: Addy Osmani**

**Main Arguments**
*   **AI coding agents tend to behave like junior engineers** by taking the shortest path to completion, often skipping critical "invisible" work such as writing specs, tests, and performing reviews [1, 2].
*   The **"Agent Skills" project provides a senior-engineer scaffolding** to ensure AI agents follow disciplined software development lifecycles (SDLC) rather than just producing unverified code [3, 4].
*   Effective agent instruction requires **"process over prose,"** meaning agents need actionable workflows with clear exit criteria rather than long essays or reference documentation [5, 6].

**Key Takeaways**
*   **A "skill" is a structured workflow**—usually a Markdown file—that defines a sequence of steps and produces verifiable evidence at each checkpoint [5, 7].
*   The framework is built on **five core principles**: process over prose, anti-rationalization tables, nonnegotiable verification, progressive disclosure, and strict scope discipline [6, 8-10].
*   **Anti-rationalization tables** are a unique feature used to counter an LLM's tendency to justify skipping steps (e.g., rebutting the claim that a task is "too simple" for a spec) [11, 12].
*   The system incorporates **Google’s engineering DNA**, including practices like Hyrum’s Law, the test pyramid, and maintaining small, reviewable PR sizes [13].

**Important Details**
*   The repo organizes **20 skills across six lifecycle phases**: Define (/spec), Plan (/plan), Build (/build), Verify (/test), Review (/review), and Ship (/ship) [14].
*   **Progressive disclosure** ensures that only relevant skills are loaded into the agent's context at any given time, preventing performance degradation from token bloat [9].
*   The skills are **portable**; they can be installed as plugins for Claude Code or used as Markdown files in tools like Cursor, Gemini CLI, and VS Code [15, 16].
*   Even without the software, teams can adopt the **"five nonnegotiables"**: surfacing assumptions, asking when requirements conflict, pushing back when warranted, preferring boring solutions, and maintaining scope discipline [17].

***

### **Get a Good Return on Your AI Investments**
**Author: Louise Corrigan (reporting on Sam Newman and Nathen Harvey)**

**Main Arguments**
*   **AI acts as an amplifier and a mirror**; it enhances efficient teams but makes friction more acute for teams with poor existing processes [18].
*   Gains in developer productivity (roughly 10% more code shipped) are often offset by **increased instability**, leading to more frequent rollbacks and fixes [19].
*   Organizations must **invest in platforms and guardrails** to manage the "verification tax" and "cognitive debt" introduced by AI-generated code [20-22].

**Key Takeaways**
*   **The "Verification Tax"**: Increasing the volume of code requires a proportional increase in the ability to verify it; otherwise, incidents will rise [20].
*   **Cognitive Debt and the "AI Vampire"**: Over-reliance on AI can lead to developers "losing the plot" of their own software, while running multiple agents simultaneously causes burnout through excessive context switching [21, 23].
*   **The Shift to Platform Engineering**: Nathen Harvey suggests that instead of just building features, more engineers should focus on **building internal platforms** that make software durable, secure, and production-ready by default [22, 24].
*   **Culture of Experimentation**: Success with AI is not about token consumption but about how many experiments a team runs and how well they distribute those lessons [25].

**Important Details**
*   DORA (DevOps Research and Assessment) data shows that **only about 24% of respondents fully trust AI output**, while 46% trust it only "somewhat" [20].
*   A new **ROI framework and calculator** has been released by DORA to help teams model real costs, including learning investments and verification overhead [26].
*   The "Beyoncé Rule" is cited as a core philosophy: **"If you liked it, you should have put a test on it"**—emphasizing that infrastructure changes alone do not catch bugs [13].
*   DORA identified **seven capabilities** that, when paired with AI, lead to better outcomes, with quality internal platforms being among the most critical [24].