skills4sh

Practical skills for developers and their agents: preserve project understanding, turn designs into code, verify interfaces, and maintain reusable workflows.

Each skill is an Agent Skills folder. The core workflows are portable. Installation paths and host-specific capabilities are documented separately; each skill tells the agent when differences in tools, permissions, or instruction loading affect execution.

Choose a skill

Start here: init-rulebook

Write your rules once. Every agent loads the ones that apply, every session.

Agent setups tend to grow by accretion: a CLAUDE.md here, a Cursor rule there, and “never push without asking” pasted into each. init-rulebook replaces that with one personal rulebook, a plain AGENTS.md index with optional rule files, kept wherever you like. Each host gets a short startup pointer; the skill does the loading.

An index can be this small:

# AGENTS.md
## Always on
- [Core](rules/00-core.md)
## Activate when code is involved
- [Code posture](rules/20-code.md)
## Action-gated: immediately before a commit
- [Commit rules](rules/02-commits.md)

Try: “Set up my personal agent rulebook.” Then, in a fresh session: “init-rulebook: status.”

Stable since 1.0.0: changes to its triggers or scope require a major version. Its scenario catalog covers all six modes, compaction, action gates, and inventory failures. See Manage your own agent rules for every mode.

More places to start

agent-memory — a shared brain for your codebase and team

The repo is the room. Git carries the context.

Preserve decisions, rationale, conventions, and handoffs as reviewable Markdown alongside the code. Commit and share that context so another developer—and their agent—can pick up the reasoning behind the work. It helps a solo developer resume a project, a team onboard a colleague, or collaborators continue across different agent tools.

Memory belongs to the project. Branches can carry exploratory thinking; review and merge establish what the team accepts. For work spanning repositories, record relevant context in each project and reference shared decisions deliberately. The skill does not automatically synchronize memory between repositories or provide per-person access controls.

Try: “Save why we chose this approach in project memory so the next developer and their agent can continue from here.”

skill-architect — make skills easier to find, follow, and verify

Turn a useful working method into a reusable skill, or audit an existing one for vague triggers, confusing steps, stale guidance, and missing checks. It helps keep the entry point concise, move detail into references, and distinguish mechanical validation from behavior that needs an actual trial.

Try: “Review this skill: would an agent know when to load it, what to do, and how to verify completion?”

localhost-screenshots — see what changed

Give interface work visible evidence. Capture a local page at requested viewport sizes, compare before and after, and investigate responsive layout problems. Use a connected Chrome workflow for quick inspection or Playwright for repeatable capture sets and visual comparisons.

Try: “Capture this local page at desktop and mobile sizes, then compare it with the screenshots from before the layout change.”

Install and try it

These are plain Agent Skills: folders with a SKILL.md entry point plus optional references and assets. Choose one skill from the table, install it, and ask your agent to use it on a relevant task.

npx skills@latest add t4sh/skills4sh --skill agent-memory --copy

Replace agent-memory with any skill name above. The interactive installer lets you choose the target agents. To browse the collection without installing:

npx skills@latest add t4sh/skills4sh --list

For a global install targeting Codex:

npx skills@latest add t4sh/skills4sh --skill agent-memory --copy -g -a codex

To try an unreleased branch, pass the ref after #. The skills CLI records it as ref in its lock entry:

npx skills@latest add 't4sh/skills4sh#' --skill agent-memory --copy -g -a codex

Use the #ref form for branch names that contain /, such as fix/my-change; a github.com/.../tree/ URL treats everything after the first / as a path. Reinstall from t4sh/skills4sh without a ref to return to released content after the branch merges.

Replace codex with a supported target such as claude-code, cursor, grok, openclaw, or github-copilot. The skills CLI supported-agents table lists current --agent values and paths. Reload or restart as required by the host.

Codex scans project .agents/skills/ from the working directory up to the repository root and user ~/.agents/skills/, and reads each skill’s optional agents/openai.yaml for its /skills display name, short description, and default prompt (checked 2026-10-20 against the Codex skills guide).

Pointer files (AGENTS.md, CLAUDE.md, Cursor rules, and similar) are documented per skill when a workflow writes them — see agent-memory agent-pointers.md. They do not change the skill’s commands.

Skill scope is different from memory scope

Installing agent-memory globally makes the workflow reusable. It does not create one global memory shared across all your projects.

Installing a skill also does not provision its dependencies. Figma workflows need the relevant Figma access and tools; screenshot workflows need a running local site and browser tooling. Check each skill’s compatibility notes for macOS, Windows, Linux, and runtime requirements.

Try the other workflows

Manage your own agent rules

Every init-rulebook mode, as introduced in Start here. Start with one ~/.agents/AGENTS.md and split it into linked rule files when conditional loading would save work.

These are requests to your agent, not shell commands. Your rules stay outside the installed skill, so updating it from t4sh/skills4sh preserves your setup. Existing ~/.agents symlinks and startup bridges can remain in place. The public t4sh/dotfiles rule set is an optional example; adopting it keeps ~/.agents as the tree and leaves its ~/.agents/rules/*.md inventory sentence unchanged.

Host compatibility: ~/.agents/AGENTS.md is this skill’s default shared rulebook location. Codex reads global instructions from its home directory (normally ~/.codex/AGENTS.md, subject to overrides); setup puts the pointer in the effective instruction file. Codex also discovers user skills in ~/.agents/skills/. The skill uses the portable SKILL.md format; each host’s startup discovery still needs configuration. See the official AGENTS.md guide and skills guide.

Structure and verification

skills//
├── SKILL.md          # Metadata, triggers, and operating instructions
├── LICENSE
├── references/       # Optional supporting documentation
└── assets/           # Optional icons, fixtures, or helper scripts

Repository checks validate skill metadata, reference links, version consistency, and file hashes. Security scans use pinned guardskills 1.2.1, with documented expected findings. These checks cover specific properties; behavioral fixtures and review provide additional evidence, and an eval scenario alone is not a recorded passing result.

See SECURITY.md for scan details, expected findings, and vulnerability reporting; .security/ contains per-skill integrity manifests. The skill authoring standard describes the review and verification requirements.

Skill files do not execute themselves. Agents may follow their instructions and run optional helpers when the task and permissions allow.

Stability: code-to-figma and figma-to-code are pre-1.0: prompt content and triggers may change incompatibly between minor releases. agent-memory, discord-harvest, eleventy-nunjucks, init-rulebook, localhost-screenshots, and skill-architect reserve breaking changes for major versions.

License

MIT — see LICENSE.