Podcast Wiki

Skill Engineering

5 min read

Summary: Agent skills are mini software: a prompt describing a workflow (what, how, when) plus resources such as scripts, CLIs, or MCP calls, a reusable process instead of a giant prompt each time. They combine deterministic code with LLM flexibility and let non-developers, especially domain experts, contribute and measure improvements. A public ecosystem of free skills and repos (launch videos, design critique, decks, research) now packages whole service jobs, so the cost is mostly model usage. In practice one orchestrator agent loading skills by scenario beats a separate agent per page or alert type. Skills are also a cost and attention tool: load them progressively, not all up front. They have a ceiling, after which teams fine-tune models. Sharing is still hard: forks diverge, external APIs change, and versioning and dependency management are immature.

Key ideas

  • Even “complex” capabilities, like letting an agent message you, can be a few lines of skill plus an API or CLI and token setup (langtalks—72-personal-assistant-agent).
  • Scripts in skills are Lego pieces the agent runs and may tweak per configuration (langtalks—72-personal-assistant-agent).
  • Sharing by copy and fork was a mistake in hindsight: bug fixes and features don’t propagate, and one user’s feature is another’s annoyance (langtalks—72-personal-assistant-agent).
  • Better: a marketplace of composable pieces, plus each user’s own workflow layered on top (“tag with meeting-tagger, but don’t change my titles”) (langtalks—72-personal-assistant-agent).
  • Org practice: require every developer to contribute to a central skills repo periodically, so the shared catalog stays the source of truth (langtalks—72-personal-assistant-agent).
  • Skills rot: a NanoClaw skill for a third-party API broke when the API changed, while old installs kept the stale skill. Skills need version control and dependency management like libraries (langtalks—72-personal-assistant-agent).
  • Managing prompts well is an unsolved challenge: in code versus a platform, with no single source of truth and syncing issues (langtalks—72-personal-assistant-agent).
  • AI-native open source: NanoClaw rejects PRs that change core code. Contributors submit skills that teach your coding agent how to modify your fork (e.g. swap WhatsApp for Telegram), so the core stays lean (langtalks—71-claw-architectures, gavriel-cohen).
  • This enables Karpathy’s “bespoke software”: everyone runs a customized version on a stable, secure base (langtalks—71-claw-architectures).
  • Agent-operable projects: a CLAUDE.md for context plus a CLI usable by the operator, by Claude in the repo, and (with approvals) by agents in containers. Setup is a script that offers to run Claude on failures (langtalks—71-claw-architectures).
  • For product makers, a skill can simply be an MD file derived from public API docs; split into several files with references if large (langtalks—69-marketing-for-agents).
  • Org-wide skills marketplace with an owner: architects or a developer per team push skills; automation can promote repo-level decisions to org rules; skills can be applied by process (e.g. frontend work) rather than by team (langtalks—68-ai-sdlc).
  • A skill packages a specialization with a very high level of detail to produce a high-quality result. Adoption is still thin in big tech: in one Apple Haifa talk, most had used Claude Code, half knew sub-agents, about three used skills (ignore-instructions—20-idan-benyon-build-ship-grow, idan-benyon).
  • Progressive disclosure is a cost and attention principle, not just tidiness: loading all skills or tasks up front wastes tokens and pulls attention toward things that won’t run (startup-for-startup—354-reliable-agents-lean-context, doron-bleiberg).
  • Skills split system prompts and speed ramp-up, but hit a ceiling; past it, fine-tune models on collected context and data, sometimes automatically at the customer site in a feedback loop (hidden-layers—dream-eran-hoffman, eran-hoffman).
  • One orchestrator plus skills beat a dedicated workflow per alert type: the investigation agent loads skills by scenario (cloud → AWS or GCP). Domain experts, not R&D, should be able to add or edit skills and measure their impact on evals (hidden-layers—vega-gili-kanfo, gili-kanfo).
  • Scoped per-page agents broke because users ask anything anywhere; one agent selecting from well-described skills improved evals immediately. A “screen awareness” skill lets users ask about what they’re looking at (hidden-layers—twine-nadav-erez, nadav-erez).
  • A skill is a set of instructions your agent reuses for a particular job, sometimes with scripts and examples: you hand it a process rather than writing a massive prompt every time. Examples now circulating: Brag (launch videos from a project or URL), PPT Master (notes → editable deck in your own template), Free Tools (Corey Haynes’ marketing collection), Grill Me (Matt Pocock), Impeccable (Paul Bakaus). Most are free; you pay for the model (startup-ideas—8-github-repos-viral-money, greg-isenberg).
  • Plain LLMs are weak at pulling insight from the social web, so tool-setup skills like Agent Reach wire up access to YouTube, Reddit and X for cross-source research (startup-ideas—8-github-repos-viral-money, greg-isenberg).
  • Third-party skills and repos can change agent settings (OpenRig does): read what an install changes before running it, and try it on a test project or a copy of your site first (startup-ideas—8-github-repos-viral-money, greg-isenberg).

Disagreements & open questions

Takeaways

personal-ai-assistants · agent-ready-codebase · ai-guardrails · agent-security · agent-ready-products

Graph View

Backlinks