--- title: "Podcast: How valuable are agent skills? Conversation with Larah Vasquez and Fabrizio Ferri-Benedetti" date: 2026-04-12 description: "Video Audio-only version Resources Topics covered in this podcast Narrative essay version of the conversation Transcript..." canonical_url: https://idratherbewriting.com/blog/ai-skills-agentic-workflows-larah-fabrizio --- # Podcast: How valuable are agent skills? Conversation with Larah Vasquez and Fabrizio Ferri-Benedetti > In this podcast, I chat with [Larah Vasquez](https://mcnuggies.dev) and [Fabrizio Ferri-Benedetti](https://passo.uno/) about using skills to extend AI capabilities, the future of agentic engineering, local models like Qwen and Gemma, and whether the tech writer role is shifting into automation architecture. We get into the memory problem in LLMs (and why some of us actually prefer the no memory to extended memory), the progression from prompt engineering to context engineering to compound engineering to orchestrating whole agent systems, and how skills are quietly forcing engineers to write down knowledge they'd never documented before. ## Video ## Audio-only version **Listen here:** [![](https://s3.us-west-1.wasabisys.com/idbwmedia.com/images/apple_podcasts.png)](https://itunes.apple.com/us/podcast/id-rather-be-writing-podcast/id277365275) [![](https://s3.us-west-1.wasabisys.com/idbwmedia.com/images/watchonyoutubeblack.png)](https://www.youtube.com/@idratherbewriting) [![](https://s3.us-west-1.wasabisys.com/idbwmedia.com/images/spotify.png)](https://open.spotify.com/show/4HeOZfPGMMfViOhVS40QBD) ## Resources - [mcnuggies - Home](https://mcnuggies.dev/) (Larah Vasquez — guest website) - [mcnuggies - Blog](https://mcnuggies.dev/blog-feed) (Larah Vasquez — guest blog) - [Passo.uno :: Technical Writing, AI & Docs Engineering :: Fabrizio Ferri-Benedetti](https://passo.uno) (Fabrizio Ferri-Benedetti — guest website) - [GitHub - armstrongl/code-docs: An LLM and human friendly docs framework for code repositories · GitHub](https://github.com/armstrongl/code-docs) (Larah Vasquez) - [Overview - Agent Skills](https://agentskills.io/home) (Agent Skills spec) - [Extend Claude with skills - Claude Code Docs](https://code.claude.com/docs/en/skills) (Anthropic) - [Gemma — Google DeepMind](https://deepmind.google/models/gemma/) (Open Source models from Google DeepMind) - [Vale: Your style, our editor](https://vale.sh/) (Vale linter) - [Larah Vasquez on LinkedIn](https://www.linkedin.com/in/larahvasquez) ## Topics covered in this podcast Here’s a list of topics we talked about. (Note: AI-generated.) - **Local LLMs and open models** — Running local LLMs like Alibaba’s Qwen and Google’s Gemma is becoming more vital to avoid API reliance, downtime, and cost — especially after Anthropic narrowed third-party API access and left OpenClaw users scrambling for alternatives. - **AI tooling strategies** — Organizations are assessing different tiers of AI tools for specific tasks to optimize performance, token usage, and compute — for example, not wasting Opus on a typo check. - **Automating style guides** — Generative AI can reverse-engineer documentation assets, such as automatically generating Vale rules from an existing style guide site instead of hand-authoring them. - **Why AI output starts bad by default** — Generic AI writing improves substantially only when grounded in high-quality signals, templates, and constraints authored by humans — the case for keeping writers in the loop. - **Model personality vs. memory** — There are real trade-offs between models that remember everything and models that come fresh to each session; sometimes you want the memory, but often you want no memory at all. - **The memory problem** — Long-term memory in LLMs is still unsolved — memory.md files get corrupted, and agents end up rereading their own sticky notes like the protagonist of *Memento*. - **Compound engineering** — A brainstorm → plan → work → compound loop where agents write lessons learned back to disk to improve over time — powerful, but extremely token-hungry. - **Token constraints and efficiency** — Running AI workflows demands careful context management; unchecked configurations pollute the context window and blow through tokens fast. - **Skills and MCP servers together** — Thin skills that call out to an MCP server for doc content tend to outperform heavy skills with everything inlined, and they’re easier to keep current because the source of truth stays in the docs. - **Subagents and worktrees** — Running skills in isolated subagents or Git worktrees lets you parallelize work and save the main context window for orchestration. - **Agentic engineering** — The progression from prompt engineering to context engineering to compound engineering to orchestrating whole systems of specialized agents working together. - **Skills as living artifacts** — Good skills get better over time by asking the agent what snagged after each run and amending the instructions — a controllable, versioned alternative to opaque model memory. - **Skills redefining tech writing** — Building skills forces product teams to write down rigorous, step-by-step knowledge they’d previously kept in their heads. - **Do more, or do the same with less?** — The honest question about AI productivity gains: shrink the team to buy more GPUs, or keep the team and finally tackle the infinite docs backlog. - **Elevating to automation architect** — The core of tech writing is shifting from authoring prose to orchestrating systems; writers who can script, configure, and curate agent workflows are becoming more indispensable, not less.