The Rookery

A field guide to how I build with AI. It connects a practical workflow for research, planning, design, implementation, shipping, maintenance, and learning with the skills and scheduled automations I use to run it.

A rookery is where corvids gather to nest. This one is named for Odin's ravens, Huginn and Muninn. Huginn is thought, Muninn is memory. Each morning they fly out to see the world, and each evening they return with what they've learned.

Guiding principles

1. Works across tools. Skills should not depend on one agent or harness.
2. Add skills sparingly. Create one only when existing tools leave a repeated gap.
3. Improve from use. Turn recurring failures into tests, rules, or clearer instructions.

Install

Install skills one at a time or all at once.

# list available skills
npx skills add jrgilbertson/the-rookery --list

# install one skill
npx skills add jrgilbertson/the-rookery --skill <name>

# or take everything
npx skills add jrgilbertson/the-rookery --all

These packages use the Agent Skills format supported by tools including Claude Code, Codex, Cursor, and Gemini CLI. Installation and runtime behavior remain harness-specific. Add -g to install once for your whole machine instead of per project.

main is the rolling catalog and the normal install source. Semantic-version Release Snapshots are immutable historical checkpoints with release notes; they are not install pins.

The workflows

Everything here fits into five core jobs, Research through Ship, plus two feedback loops: Maintain and Learn. The walkthroughs in WORKFLOWS.md are how I run that loop. The skills in this catalog plug into it and work in other editors and harnesses too.

• Research. Gather current evidence, test it against opposing views, and curate enough context to plan.
• Plan. Define what to build, the guardrails execution has to honor, and how success will be verified.
• Design. Set the visual direction and written design system for interface work.
• Build. Implement the plan in bounded, independently verified slices.
• Ship. Review and verify the finished change before merging and documenting it.
• Maintain. Keep repositories healthy by turning recurring problems into durable safeguards.
• Learn. Turn experience into linked knowledge and new questions for Research.

Testing with agents explains how I give agents room to work while permissions, end-to-end tests, and independent checks protect the result. Use it when setting up your own environment and deciding where safeguards belong. For skill validation, SKILLS.md distinguishes deterministic repository checks from behavioral evaluation through the evals plugin's evals-start workflow. creating-portable-skills is the first dogfood target; the other skill packages await revalidation after the eval reset.

The skills

• checking-simplicity. Find opportunities to safely simplify a design, plan, or approach against the current need. Use it when you ask to simplify something, including during a build, or when the next step adds durable machinery that isn't tied to that need.
• creating-portable-skills. Create, revise, migrate, or audit a portable skill package. Use it to deliver a self-contained package; this repository routes behavioral evaluation through the evals plugin's evals-start workflow.
• managing-issues. Manage GitHub or Linear issues and their native parent and blocker relationships through the repository's canonical tracker. Use it to read, draft, create, or update a single issue or multi-issue graph, and to assess readiness, dependencies, or completion.
• managing-personal-crm. Keep relationship context in one note per person while messages and other raw interactions stay in their original apps. Use it to prepare for someone, capture an interaction, reconnect an overdue relationship, or find who could help with current work.
• personal-chief-of-staff. Turn information from your configured sources into a daily, weekly, or quarterly review. Run it when you need to orient, reflect, and decide what to do next across several parts of your life and work.
• repo-gardener. Sense one repository, dispatch Executors that each ship one reviewable pull request through the installed Compound Engineering shipping skills, and post one morning report. Run it on a schedule by name; a human still merges.
• reviewing-meetings. Turn completed meetings into draft notes and follow-up actions grounded in the meeting source. Run it after a meeting or during a catch-up when you want to capture outcomes and approve each follow-up before anything is written.

The automations

I run recurring reviews in Orca to find improvements in my development work. The automation catalog contains copyable prompts, schedules, prerequisites, and setup instructions. Configure these jobs in Orca separately from installing skills.

• Sessions audit. Review recent AI sessions for recurring friction and improvements to code, tools, and guidance.
• Test audit. Review the tests, find gaps in what they protect, and recommend improvements.
• Optimize CI. Review continuous integration and recommend improvements.
• Weekly codebase design. Review codebase design and recommend improvements.
• Coding standards audit. Review adopted coding guidance against repository evidence and current primary sources.

My other projects

• Networked Thinking. My system for turning what I learn into durable, linked notes, explained through a book and website. Use it when something I learn should become knowledge I can connect and reuse.
• Networked Thinking Skills. A collection of skills and deterministic helpers for Networked Thinking notes and tasks in Obsidian. Use it when you want an agent to create or audit atomic notes or manage Obsidian tasks.

Standing on

This system builds on work by people who share theirs. Use them directly.

• Compound Engineering by Trevin Chow (@trevin) and Kieran Klaassen (@kieranklaassen). Plans, implements, reviews, and documents the development work.
• Matt Pocock's skills by Matt Pocock. The targeted grilling and issue-planning patterns that shaped managing-issues, plus writing-for-agents for documents that an agent consumes.
• humanize by Harshaneel Gokhale. Its ai-check skill flags AI-sounding writing.
• Impeccable by Paul Bakaus (@pbakaus). Shapes design direction, critiques interfaces, and keeps design systems coherent.
• last30days by Matt Van Horn (@mvanhorn). Recent-signal research across Reddit, X, YouTube, HN, and the web.
• Orca by Jinjing Liang (@JinjingLiang). The agentic IDE I use to run this workflow across several harnesses. The skills also work in VS Code, Claude Cowork, Codex, and other compatible tools.

Contributing

Fixes and portability PRs are welcome. New skills start as an issue. See CONTRIBUTING for how this repo stays healthy.

This is a solo-maintained project with no support SLA. Please report security issues privately through the process in SECURITY.md, and use the Code of Conduct for community participation.

License

MIT.