jaktestowac/authoring

Audits a collection of skills, agents, commands, and rules as one system: inventory, conformance to the house conventions, description trigger quality, overlap between near neighbours, broken cross-references, orphaned resources, drift between sources and generated copies, and which assets have no behavioural test. Use when preparing a release of the collection, after a batch of new assets, or when assets start colliding or triggering on the wrong requests. Read-only: it reports and never fixes.

Creates a new agent skill (a SKILL.md folder with supporting resources) for GitHub Copilot, Claude Code and OpenAI Codex, either as a contribution to this collection or in the user's own project, and validates it. Use when you want a guided, step-by-step build of one skill; for the underlying method, see the creating-skills skill.

Creates custom agents (subagents): a focused role with its own instructions, the fewest tools the job needs, explicit boundaries, stop conditions and an output contract, written as a Copilot .agent.md, a Claude Code agent file, a Codex agent TOML, or this collection's single-source agent format. Use when a recurring job needs its own role or tool limits, when an existing agent has grown too broad or has too many tools, or when a review step should run separately from the work it reviews. For several cooperating agents with an orchestrator, use creating-orchestration-packs; for reusable procedure without a role, use creating-skills.

Creates lifecycle hooks: small scripts the host runs at session start, before or after a tool call, on prompt submit or at stop, to log, add context, warn or block deterministically. Covers GitHub Copilot hooks.json, Claude Code hooks in settings or a plugin, OpenAI Codex .codex/hooks.json, and this collection's per-host hook source. Use when a guardrail must hold regardless of what the model decides (block a dangerous command such as git push or rm -rf, check a file after every edit, inject context at start), when porting a hook between tools, or when a hook is slow, noisy or blocks the wrong things. For guidance the agent may weigh, use creating-instructions.

Creates instruction files and scoped rules that agents apply automatically: repository-wide guidance (copilot-instructions.md, CLAUDE.md, AGENTS.md) and file-scoped rules (Copilot .instructions.md with applyTo, Claude Code .claude/rules/ with paths, Codex AGENTS.md), plus rules in this collection's single-source format. Use when encoding project conventions, choosing how narrowly a rule should apply, splitting an overgrown instructions file, or porting rules from one tool to another. For step-by-step workflows, use creating-skills; for enforcement that must not depend on the model, use creating-hooks.

Creates orchestration packs: an orchestrator plus a few narrow subagents (explore, plan, implement, review, verify) with minimal tools per role and one shared handoff-packet contract, built so they run in GitHub Copilot (orchestrator agent with handoffs) and in Claude Code and OpenAI Codex (orchestrator as a command in the main conversation, since subagents there cannot start subagents). Covers this collection's src/packs/ format and hand-written packs in a user's project. Use when one agent's role has grown too broad, when phases need different tools or a fresh context, when review must be independent of implementation, or when a pack fails validation because a handoff target does not resolve. For a single agent, use creating-custom-agents.

Packages skills, commands, agents, hooks and MCP servers as installable plugins with a marketplace for GitHub Copilot, Claude Code and OpenAI Codex: choosing what to bundle, writing each host's manifest and marketplace entry, and checking the result installs. In this collection, plugins are generated from bundles.yaml by npm run build. Use when a skill or group of assets should install as a unit, when adding or reshaping a bundle here, when publishing a plugin from your own repository, or when CI reports generated plugin files out of date. For the skill being packaged, use creating-skills first.

Creates reusable commands the user starts on purpose (/name, or $name in Codex): one job, declared inputs with hints, ask-when-missing behaviour, a concrete output and location, a verification step and guardrails. Writes them as a Copilot .prompt.md, a Claude Code user-run skill or command, a Codex skill with implicit invocation turned off, or a command in this collection's source format. Use when a team keeps re-typing the same request, when a workflow needs a fixed entry point that routes to an agent, or when an existing prompt produces vague output. For know-how the agent should pick up on its own, use creating-skills.

Creates agent skills (a SKILL.md folder plus optional resources) that load in GitHub Copilot, Claude Code and OpenAI Codex: a description that triggers on the right requests, a lean body with progressive disclosure, checkable steps, and a validation pass. Use when packaging repeatable know-how as a skill the agent picks up on its own, splitting a bloated SKILL.md into body plus resources, fixing a skill that fires on the wrong requests or never fires, or adding a skill to this collection, even if the user only says 'turn this into something reusable'. For a command the user runs on purpose, use creating-prompts; to test how an existing skill behaves, use testing-agent-skills.