
Antigravity separates instructions that should load only when relevant from guidance that should apply on every task. A skill is a folder containing a SKILL.md file, with optional extras such as a scripts folder, that adds on-demand instructions to the main agent's conversation using the same tools and model. Because a skill loads only when the agent decides it matches a request, telling skills apart from rules, custom subagents, and hooks is a central detail for the Professional Agentic Architect exam.
A skill is a folder with a SKILL.md file inside it, rather than a single Markdown file. The instructions belong in SKILL.md, and optional supporting files can sit in subfolders such as scripts.
The frontmatter in SKILL.md requires both a name and a description. At the start of a conversation, Antigravity shows the agent only each skill's name and description, and the agent reads the full instructions only when that description makes the skill look relevant to the task. The description is therefore the main lever on whether a skill gets used, and it works best when it states both what the skill does and when it applies.
This on-demand loading is what makes skills economical. A skill's instructions stay out of context until a task calls for them, unlike an always-on rule, which loads on every turn whether or not the current request needs it.
Antigravity reads skills from two locations depending on whether the skill should be shared with a team or kept personal across workspaces:
.agents/skills/<name>/SKILL.md inside a repository. Committing the folder shares the skill with everyone who opens the repository. Antigravity still reads the older .agent/skills/ path, though .agents/ is the current default.~/.gemini/config/skills/ on the developer's machine. Skills there are available in every workspace the developer opens in the Antigravity IDE without committing anything to a repository. Antigravity still supports the older ~/.gemini/antigravity/skills/ location.Skills sit alongside rule files, which live in rules folders at the same two levels, .agents/rules/ in a workspace and the rules folder in the global config directory. Antigravity treats the files in those folders as rules rather than skills. A developer's personal GEMINI.md in ~/.gemini holds that developer's global rules.
A skill loads on demand when the agent judges it relevant, so the agent can pass over it. Standing instructions work differently, and Antigravity offers two workspace mechanisms that apply instructions without relying on the agent's judgment:
GEMINI.md and AGENTS.md files in project directories as rules that are always active within their directory. These files take no frontmatter, and committing one at the repository root shares it with everyone who opens the repository..agents/rules/ with its trigger set to always_on applies on every turn in that workspace. Rules with other triggers load only under certain conditions, such as manual, which loads a rule when someone mentions it.On the Professional Agentic Architect exam, use a skill for specialized procedures that apply only to certain tasks and should stay out of context otherwise. Use GEMINI.md, AGENTS.md, or an always_on rule when standing guidance must apply unconditionally.
A skill adds instructions to the main agent's own conversation, so the work runs with the same conversation history, tools, and model. Custom subagents are the mechanism for handing work to a separate agent. They are defined as Markdown files under .agents/agents/, and committing a file shares the subagent with everyone who opens the repository. Their configuration covers several things a skill cannot change:
tools list that limits the subagent to the tools it names.model setting, such as pro, to pick the model tier.description field tells the main agent when to delegate a task to it.Antigravity also has a built-in self subagent, which is a clone of the calling agent with identical toolsets.
Skills and hooks differ in who decides when they run. A skill runs when the agent decides to use it. Hooks run automatically at fixed points, without the agent or a developer choosing them. Each hook event corresponds to a different point in the agent's loop:
SessionStart: Fires once when a new conversation starts, before the first model call. Its output can include injectSteps, which add steps to the conversation at the start of the session.PreInvocation: Fires before every model call.PreToolUse: Fires before a tool executes and returns a decision about that single tool call.PostToolUse: Fires after a tool call completes, and its output is an empty object.Stop: Fires when the execution loop ends.Our Professional Agentic Architect practice exams test Antigravity skills alongside rules and custom subagents, with detailed explanations that drill these distinctions.