Use the method behind the work
Barnaby Robson's public skills collection contains eight curated skills for people who use agents to research, analyse and write. Each skill gives an agent instructions for a repeatable job.
Read the source on GitHub. Get individual ZIP files, the complete collection and installation guidance from the downloads page.
Give shared instructions one home
The Home Repo keeps the root AGENTS.md as a map of repositories, projects and vaults. Its home/global.md holds universal rules. Claude Code, Codex and Gemini can each link their global instruction file to that same contract.
Shared skills live in skills/<name>/ and mount into ~/.claude/skills/ and ~/.codex/skills/. A project's own skills have one canonical source at .agents/skills/<name>/SKILL.md. Relative links let other runtimes read that source without maintaining separate copies.
Keep project rules in the project's AGENTS.md. Keep private paths, credentials and machine settings outside the public collection.
The eight published skills
vault-setup
Creates personal, project, operations, wiki and agents vaults. Each starts with a folder tree, templates and its own instruction file.
recall
Finds past work in session notes. It reads native transcripts when notes are missing or more detail is needed.
retrospective
Reviews a session or a bounded sample of recent sessions. It verifies repeated problems and proposes small fixes. It uses recall for historical evidence.
orchestrator
Coordinates T3 Code, cmux and CLI subagents. It routes each job by model and effort, manages isolated work and writes handoff briefs at 30% context usage. Its helper script requires PyYAML. Configure its routing table before dispatching workers.
cmux
An optional standalone terminal host driver. It controls cmux workspaces and browser panes, and calls orchestrator for routing policy. Install both skills when using cmux dispatch.
graph
Turns a YAML dependency work graph into an interactive page. It supports dependency tracing, foldable branches and node inspection. Content hashes mark outputs stale when their inputs change.
editorial-review
Checks prose and argument before a reader sees the work. It reviews structure, evidence, clarity and signs of machine-written prose.
skill-manager
Writes and reviews skills. It includes description rules, a checklist, a validator and a whole-tree audit.
Check web platform choices with Modern Web Guidance
Modern Web Guidance gives coding agents targeted advice on browser APIs, accessibility and performance. Chrome, Edge, Firefox and MDN steward the project with the web development community.
Run npx modern-web-guidance@latest to use its command line. Search for a specific task, then retrieve the matching guide. This progressive disclosure keeps the full guidance library out of the agent's working context.
The source repository holds the authoring and evaluation tools. Nightly evaluations compare agents with and without guidance across Claude, Codex and Gemini. The results help maintainers refine advice and prune material that models already know.
Check each recommendation against your browser requirements. For this site, Safari and iOS Safari must retain working behaviour. Experimental browser APIs need a separate decision.
Put skills inside a working harness
Start with building skills and workflows. Read The Harness for the wider system of instructions, context, tools and checks around the model.
Use Using GitHub to learn how to develop, review and release skills in a shared repository.
On pstack traces Lauren Tan’s skills and the mental models behind them. It shows how verification, shared constraints and repeatable workflows support agent work.