AI agent skills that execute the same way every time.
Fix it once, everywhere.
AI agent skills are packaged instructions an agent loads when a task matches: a SKILL.md file with a name, a description and a workflow, as defined by the open Agent Skills format. Vodou runs that format natively and adds what the format leaves to the model — an inline AGENT_ACTIONS block that Vodou's workflow driver executes step by step, so tool calls run in order, stopping points wait for you, and no step is skipped or improvised. A skill library is active on install, every skill's trigger phrases route plain-English requests to it, more install from a sha256-pinned catalog, and Skill Consoles let an agent mint a new skill at runtime with its own tab and schedule.
- Vodou runs unlimited AI agent skills — every one you install or write — each a SKILL.md file with YAML frontmatter, routed by its trigger phrases before any model call.
- Vodou skills carry an inline AGENT_ACTIONS block — server, tool, args, loops, captured variables — that the workflow driver executes deterministically, so the model cannot skip a step.
- Vodou installs skills from the VodouAI/vodou-skills-catalog with skill install, verifies each one's sha256, and merges upstream changes into your forks with skill diff and skill update.
- Vodou Skill Consoles are skills an LLM creates at runtime: each gets its own chat tab, an optional cron schedule, and /refine, /cron and /disable commands.
What are AI agent skills?
AI agent skills are reusable packages of instructions — and optionally scripts and resources — that an AI agent loads when a task matches the skill's description. The open Agent Skills format (agentskills.io) defines a skill as a folder with a SKILL.md file: YAML frontmatter with a name and description, then Markdown steps. Claude Code, Codex, GitHub Copilot and Vodou all read that file.
There are three ways to extend an agent, and they are easy to confuse. Rules (Cursor's .mdc files, AGENTS.md, CLAUDE.md) are always-on text loaded into every conversation. Skills are on-demand: the agent sees a one-line description of each and loads the full file only when a task matches. MCP servers are running processes that expose typed tools. A skill usually uses MCP tools; it does not replace them. Vodou has all three — skill routes, any MCP server you add, and hooks that write rules files for Claude Code and Cursor — and this page is about the middle one.
A Vodou skill is a folder under skills/ with a SKILL.md whose frontmatter carries name, description, version, trigger_phrases, stopping_points and actions, plus optional required_tools, allowed_tools, disallowed_tools and when_to_use; vodou-core skill validate checks every file against a JSON schema. Vodou runs on your machine and needs a free Vodou account so the engine can be licensed to you. The skills, the gateway that runs them and the MCP servers are Apache-2.0 in the VodouAI/OS repository; the Rust engine that routes to them is proprietary.
Why do AI agent skills matter?
Because without them you re-teach the procedure every time — the seven steps of the weekly report, the order the tools go in, the check at the end — and with instruction-only skills the model still decides whether to follow step three. Vodou's skills fix a workflow once and run it the same way in the Console, in Slack, on a schedule or on the kanban board, with the steps you wrote in the order you wrote them.
Re-explaining the procedure
A prompt is forgotten when the chat ends; a skill is a file with a name that a plain-English phrase brings back.
Advice is not execution
In Claude Code, Codex and Copilot a skill is text the model reads and follows; it can reorder, merge or skip steps, and a “send” step can fire before the check that was supposed to precede it. Vodou's driver runs the steps; the model summarizes the results.
Copies drift
Skills pasted between machines carry no version and no hash. Vodou's catalog pins a sha256 per skill and merges upstream updates into your fork with conflict markers instead of overwriting your edits.
How does Vodou run AI agent skills?
Vodou matches your request to a skill before any model call, loads its SKILL.md, executes the skill's AGENT_ACTIONS steps through the workflow driver, and stops at every stopping point until you choose.
- STEP 1
Route before inference.
Each skill's trigger_phrases become rows in Vodou's intent table (priority 40; rows you curate at 80 or above are never touched), so “calendar quick status” reaches the calendar skill deterministically, without a model guessing. A when_to_use line lets paraphrases match too. vodou-core intent-signal "<prompt>" shows which skill would fire and why.
- STEP 2
Execute the steps, not a paraphrase of them.
The SKILL.md body carries an inline <!-- AGENT_ACTIONS: {...} --> block — the source of truth; actions.json beside it is a generated cache. Each step names a server, a tool and args with {{VAR}} templating; loop repeats a step with {{i}}, capture lifts a field from one step's output into the next, stream_progress shows long steps live. A tool error aborts the chain and is shown inline; on success the model writes the summary in the skill's voice.
- STEP 3
Stop where the skill says stop.
Stopping points are numbered menus (1–9). A skill declared stopping_points: required cannot finish without reaching one; the reply stays inside the skill, its state persists for 30 minutes and survives a gateway restart. In workflows, any step that sends, posts or deletes is held behind an ask me gate that reaches you in the Console, the browser side panel or any messaging channel you use.
- STEP 4
Install, fork and update without losing your edits.
vodou-core skill install vodou.<id> fetches from the public VodouAI/vodou-skills-catalog, verifies the sha256 over SKILL.md plus actions.json, and registers the triggers. skill fork snapshots the upstream baseline, skill diff shows your changes, and skill update performs a three-way merge — conflicts get markers, never silent overwrites. skill import <path-or-url> adapts Claude Code commands and agents, Hermes skills and raw Markdown into the same shape.
What can you do with Vodou's skills?
An agent skills library on your own disk
Every skill you install or write, in three tabs at localhost:8765/#/capabilities?tab=skills: SubAgent personas by department, Workflows by source (core, catalog, imported, forked, community), and My Skills. Toggle, run, edit, or open the visual builder for a skill's actions.
SKILL.md files you can share by email
The single file carries the frontmatter, the steps and the stopping points; git show it and you have everything. skill validate checks the schema, skill audit reports the whole folder's health in one pass.
Skill Consoles — skills created while you work
Any connected LLM (or the sidebar wizard's Draft with AI) calls vc_skills_create and Vodou mints a chat tab named after the skill, routes messages there through its prompt template ({{user_message}}, {{history}}, {{invoke_tool:…}}, {{invoke_recall:…}}), and takes /refine, /cron, /disable, /clone, /snapshot, /history and /model in the tab. Twenty can be active at once by default.
Agent skills vs MCP, settled
In Vodou a skill calls any MCP server you connect. required_tools declares what it needs; allowed_tools and disallowed_tools scope what an agent running the skill may touch, and become --allowed-tools flags when a board worker runs it.
Skills on a schedule and inside workflows
Describe a recurring job in a sentence, see the plan card, press Save + schedule: Vodou writes the skill, its trigger phrase and its scheduled task, and reports each by name. Independent steps run in a together: block; an automation can run a skill as its action.
Bring skills from Claude Code and Hermes
skill import detects the format (Vodou-native, Hermes, Claude Code command, Claude Code agent, raw Markdown, URL) and lands it under skills/imported/ with imported_from provenance and the upstream sha, so you can tell later whether an upgrade is safe.
Vodou lets me focus on the outcome instead of figuring out which model, tool, or workflow to use. It removes a lot of the manual coordination from AI-assisted work.
Sources Agent Skills specification · Claude Code skills documentation · Vodou docs: skills.md · Vodou docs: skill-console.md · VodouAI/vodou-skills-catalog
How does Vodou compare to the Agent Skills format, Claude Code skills and Cursor rules?
The Agent Skills format is supported by 46 listed clients, including Claude Code, Codex, GitHub Copilot and Cursor, and skills.sh alone counts 1.4 million installs — a shared ecosystem far larger than Vodou's own catalog. Cursor rules are context files, not skills. None of them executes a skill's steps deterministically: the model reads instructions and decides. Vodou uses the same file shape, adds executable AGENT_ACTIONS, enforced stopping points, a hash-pinned catalog and runtime-created Skill Consoles — but only inside Vodou.
| CAPABILITY | VODOU SKILLS runtime | AGENT SKILLS FORMAT agentskills.io | CLAUDE CODE SKILLS Anthropic | CURSOR RULES Cursor | SKILLS.SH / AGENTICSKILLS directories |
|---|---|---|---|---|---|
| What it is | Runtime that routes to, executes and schedules SKILL.md skills | Open specification (Anthropic-originated) | Skills feature of one agent, built on the open format | Context files for one IDE (Cursor also loads SKILL.md skills separately) | Directories of SKILL.md repos |
| File format | Same shape as the spec: SKILL.md + YAML frontmatter with name, description, plus top-level version, trigger_phrases, stopping_points, actions — which skills-ref validate rejects as unexpected fields | SKILL.md; name ≤64 chars, description ≤1024; optional license, compatibility (≤500), metadata, allowed-tools; validator rejects other top-level keys | SKILL.md; spec fields plus when_to_use, disable-model-invocation, context: fork, model, hooks, paths and more | .cursor/rules/*.mdc (plain .md ignored) or AGENTS.md | Whatever the repo ships |
| Who executes the steps | Vodou's workflow driver runs server/tool/args steps with loops and captures; the model summarizes | The agent; “no format restrictions” on the body | The model; !`command` runs shell before the model reads the skill | The model; “rule contents are included at the start of the model context” | The host agent |
| Stopping points / approvals | Numbered menus enforced; ask me gates in workflows | None in the spec | None enforced; disable-model-invocation makes a skill manual-only | None | None |
| How a skill is chosen | Every skill's trigger phrases matched before inference; when_to_use for paraphrases | Description matching by the agent | description / when_to_use matching, or /skill-name | Always / Apply Intelligently / file glob / Manual (@-mention) | Host agent's matching |
| Install and update | skill install from a sha256-pinned catalog; fork, diff, three-way update | skills-ref validate; no registry | Personal or project folders, plugins, managed settings | git; Team Rules dashboard (Team and Enterprise plans) | npx skills add owner/repo, npx skills update; audits by Gen, Socket and Snyk; no hash or lockfile pinning documented |
| Runs inside other agents | Vodou only (imports from Claude Code and Hermes) | 46 clients listed, incl. Claude Code, Codex, Copilot, Cursor, Gemini CLI | Claude Code; the same files load in other Agent Skills clients | Cursor | 20 agents on skills.sh; the CLI lists 79 |
| Ecosystem size | Unlimited — install or write your own | 46 supporting clients listed | Bundled + community | — | 1.42 M installs (skills.sh); 193+ curated (AgenticSkills) |
| Skills created at runtime | Skill Consoles with a tab and cron | No | No | No | No |
| Memory inside a skill | {{invoke_recall}}; {@my team} fills personal facts; workers start with memory | No | Via CLAUDE.md / auto memory | Rules are the memory | No |
| Code you can read | Skills, gateway, MCP servers Apache-2.0; engine proprietary | Spec is open | Claude Code closed; your skills are yours | Closed | skills CLI MIT; listed skills open on GitHub |
Frequently asked questions about AI agent skills
What are agent skills in AI?
Agent skills are reusable packages of instructions, and optionally scripts and resources, that an AI agent loads when a task matches the skill's description — a folder with a SKILL.md file, per the open Agent Skills format. Vodou runs as many as you install or write, routes plain-English requests to them with each skill's trigger phrases, and executes the steps each skill declares in its AGENT_ACTIONS block instead of leaving the order to the model.
What is a SKILL.md file?
A SKILL.md file is the Markdown file at the root of a skill folder: YAML frontmatter with at least a name and a description, followed by the instructions. Claude Code, Codex and Vodou all read it. In Vodou the same file also carries trigger_phrases, stopping_points and an inline AGENT_ACTIONS block, so one file you can email or git show holds the routing, the steps and the menus.
What is the difference between agent skills and MCP?
Skills are instructions an agent loads on demand; MCP servers are running processes that expose typed tools. A skill tells the agent how to do a job; an MCP server gives it something to do the job with. In Vodou a skill's steps call any connected MCP server by server and tool name, required_tools declares what it needs, and allowed_tools limits what a worker running it may touch.
What are agent skills vs tools?
A tool is one callable action — search a calendar, send a Slack message — usually exposed by an MCP server. A skill is a workflow that strings tools together with the judgement, checks and stopping points around them. Vodou keeps the two distinct: tool routes and skill routes side by side in one intent table, and a skill's AGENT_ACTIONS names the exact tools it runs.
Do agent skills work in Claude Code?
Yes — Claude Code reads SKILL.md files from ~/.claude/skills/ and .claude/skills/ and invokes them by description or /skill-name. Vodou complements that: skill import adapts Claude Code commands and agents into Vodou skills, and Vodou's Claude Code hooks inject your memory on session start and every prompt, so a skill in either place runs with the same context.
Is there an agent skills standard?
Yes. The Agent Skills specification at agentskills.io, originated by Anthropic, defines the SKILL.md format — required name and description, optional license, compatibility, metadata and allowed-tools, plus scripts/, references/ and assets/ folders — and OpenAI's Codex, Microsoft's Agent Framework and Google Cloud document the same standard. Vodou's skills follow that shape and add fields for routing and execution.
Where can I find an agent skills library?
Public directories such as skills.sh (Vercel, 20 agents, npx skills add) and AgenticSkills (193+ curated skills) list SKILL.md repos for Claude Code, Codex, Cursor and others. Vodou ships a skill library you can extend on install and a public catalog, VodouAI/vodou-skills-catalog, where every entry carries a sha256 that skill install verifies before writing the skill to disk.
How do I install agent skills?
In Vodou, run vodou-core skill install vodou.<id> for a catalog skill or skill install <folder> for a local one; the catalog copy is sha256-verified and its trigger phrases are registered automatically. skill import <path-or-url> brings in Claude Code, Hermes or plain-Markdown skills. The Skills page in the Console does the same with a Browse Catalog modal and an Import button.
What does Vodou mean by a skill gateway?
One place through which every skill is found, run and updated, whatever it does. In Vodou that is the engine's intent table (trigger phrases that route to skills and to tools), the workflow driver that executes AGENT_ACTIONS, and the catalog commands that install and update skills — reachable from the Console, the browser side panel, any messaging channel, the command line and any MCP client. It is not an industry term; this page is about AI agent skills.
More from Vodou
- AI memoryMemory you own that compounds with every session.
- AI harnessPlain English in, executed answer out.
- Agent harnessSkills, MCP, workers and hooks — already wired.
- Browser extensionYour chats, captured locally and typed back anywhere.
- AI orchestrationRoute before inference. Run tools together.
- AutomationAgents that run unattended, within budgets you set.
- MCP gatewayOne MCP host for every editor on your machine.
- Local AI agentRuntime, memory and tools on your machine.
- Claude Code memoryMemory in Claude Code, Cursor and Codex on every prompt.
Fix it once, everywhere.
Write the workflow once as a skill — or let Vodou write it from a sentence — and it runs the same way in the Console, in Slack, on a schedule and on the board. Free for 60 days, no credit card.
Updated