The coding agent is turning into a package manager — for expertise, not code.

A CLI plus Claude Code skill that converts a technical book — PDF, EPUB, or DOCX — into a compact, navigable Agent Skill instead of a wall of PDF text. It extracts the book into a ~4K-token SKILL.md of mental models plus per-chapter files (~1K tokens each) that load only when you ask for them, so you query /designing-data-intensive-apps replication and it pulls the relevant chapter rather than the whole volume. Extraction runs locally (Docling for tables/code, pdftotext for prose); the analysis pass runs through the agent you already have, so there's no separate API bill. Install as a skill with git clone … ~/.claude/skills/book-to-skill, or as a tool with pip install "book-to-skill[pdf,epub,docx]", then /book-to-skill ~/my-book.pdf ddia. 7.3k stars, MIT.

Reach for it when you keep pasting the same reference book into context and paying the token tax every turn — the project claims 24–51x savings versus dumping the full text. Delete the ritual of dropping whole chapters into the chat window and re-uploading them tomorrow. The tradeoff is two-sided: chapter auto-detection needs explicit "Chapter N" headings, so a book with fancy section titles degrades; and a skill is a lossy compression of the book, not the book — it's a fast index to what you already trust, not a substitute for having read it.

A Claude Code slash command (/council, also a Codex skill) that convenes 18 named personas — Aristotle, Feynman, Kahneman, Torvalds and others — to deliberate a hard decision across multiple rounds: independent analysis, cross-examination, then a verdict. The mechanism is the point: dissent quotas and "novelty gates" force genuine disagreement instead of a chorus, and the output surfaces unresolved questions rather than manufacturing consensus. It spreads members across whatever providers you have wired — Claude, OpenAI, Gemini, Ollama, NVIDIA NIM — so the diversity is across models, not just prompts. Install with ./install.sh, then /council Should we open-source our framework?. 2.4k stars, MIT.

Reach for it on a genuinely ambiguous, high-stakes call where a single model's confident one-shot answer is itself the failure mode — you want the argument, not the verdict. Delete the "ask the model once, get a plausible paragraph, move on" habit for decisions that deserve to be contested. The honest tradeoff: full mode fans out to many model calls, and pointing it at several paid providers multiplies the bill fast — Quick and Duo modes exist for exactly this, and the personas are a framing device for diverse priors, not actually Feynman. Use it for the handful of decisions that earn the spend, not your daily standup.

Google's official CLI and seven skills that turn any coding assistant into an ADK agent-builder: scaffold a project, write the agent, evaluate it, deploy it. Install just the skills with npx skills add google/agents-cli, or the full CLI with uvx google-agents-cli setup. The reason it belongs in today's slate is the delivery mechanism as much as the content — a first-party team at Google decided the way to hand you its framework expertise is skills add, the same socket the two picks above use, not a documentation site you tab away to. 4.5k stars, Apache-2.0, created April 2026.

Reach for it when you're building on Google's Agent Development Kit and keep re-deriving the scaffolding from sample repos — the skill carries the current pattern straight into your agent. Delete the docs-and-samples tab-hunt for ADK boilerplate. The tradeoff is two locks you should know before adopting. First, cloud: the create, run, and eval skills work locally with an AI Studio key, but deploy, publish, and observability require a Google Cloud project — the free part is local development, the production part is GCP. Second, framework: this teaches ADK specifically, so it's Google's agent stack you're installing expertise in, not a neutral one. Great if that's your bet; it's not a portable skill.

Off the thread but worth knowing: agentOS (github.com/rivet-dev/agentos, Apache-2.0, 3.4k stars) is a portable "operating system for agents" — an in-process kernel with a virtual filesystem, process table, and network stack that runs a coding agent's code inside your app's own memory instead of booting a VM. The benchmarks are the pitch: ~4.8ms cold starts versus ~440ms, and roughly 8x less memory, self-hosted as an npm package (or hosted via Rivet Cloud). It's off today's thread because it's about where the agent runs, not what you load into it — but if sandbox cold-start latency or cost is your bottleneck, it's the layer to watch. Created 2024, actively shipping (v0.2.4) — established and re-trending, not new this week.