Skill canvas

A procedure skill is a skill you build as a flowchart. You lay out the steps on a canvas, and Studio compiles them into the numbered instructions the agent reads. Use it when the order of steps, the branches, or the checks between them matter. For a skill that is just guidance, a plain markdown skill is simpler.

The flowchart is saved with the skill, and the instructions are generated from it. You can’t edit them by hand: the skill’s Edit form shows an Edit on canvas link in place of the content box.

Open the canvas

Only new skills can use the canvas. A skill written as markdown can’t be converted to one, and saving a graph onto it is refused.

Go to Skills & Tools → Skills and click New skill. You need the editor role.
In the Add a skill panel, choose Procedure canvas instead of Write markdown.
Fill in Name (lowercase letters, digits and hyphens), Display name, Description, Trigger phrases and Tags. Write the description well: the router picks skills by it.
Click Create and open canvas.

Later, open the skill’s detail page and click Edit on canvas. A canvas skill left before its first save shows Build on canvas instead.

The canvas needs a wide screen; on a phone it says so.

The layout

The top bar has a Skills back button, the skill name, and these controls:

Control	What it does
Agent picker (Check against agent)	Picks the agent the checks and Try it use. No agent skips the agent-specific checks.
Draft from description	Has the platform LLM draft a graph from a description.
Auto-arrange	Lays the steps out again.
Compiled / Hide compiled	Shows the compiled instructions in a drawer under the canvas.
Save	Saves the graph. Off until you change something.

Left column, Add a step. One button per step type.
Middle, the canvas. Drag steps, connect them by dragging from one to the next, select a step or an edge to change it.
Right column. The inspector for the selected step, then Checks, Bundle files and Try it.

A new canvas skill starts with one Step joined to one End, and a new step lands below the selected one. The start step is marked start; select another step and click Make this the start step to move it. While other steps exist, the start step can’t be deleted. Leaving the page with unsaved changes asks first.

Draft from description

Click Draft from description, describe the procedure (at least 20 characters), and click Draft. The platform LLM returns a graph, which is laid out and put on the canvas. If the canvas already has work, you’re asked before it is replaced. Nothing is saved until you click Save.

It needs the platform LLM to be configured. If it isn’t, the dialog says “The platform LLM is not configured.”

The eight step types

Step	What you fill in	What the agent is told
Step	Instruction	Do this.
Decision	Condition	Choose a branch. Each outgoing edge needs a label.
Tool	Tool, What to do with it	Use this tool and wait for its result.
Ask the user	Question, optional Options (one per line)	Ask, end the turn, and wait for the answer before going on.
Verify	Check that	Actually perform the check. A fail edge says where to go if it doesn’t hold.
File	Bundle file, optional Note	Read, fill in, run or use a file from the skill’s bundle.
Delegate	Kind (`skill`, `expert` or `subagent`), Target, optional Note	Invoke that skill, consult that expert, or spawn that sub-agent, and wait for the result.
End	What to report	Finish. Mark every step not taken as skipped, then stop.

Names in Tool, Bundle file and Target are free text. The box suggests names: built-in tools plus the chosen agent’s MCP tools, the bundle’s files, and the chosen agent’s skills, experts or sub-agent types.

Edges and labels

Only a Decision can have several outgoing edges. Every other step has at most one, plus one fail edge on a Verify.
The first two edges from a Decision are labelled yes and no. Select an edge to rename it, for example to branch on several answers. A condition written as a question works best with several labelled edges.
A Verify’s second edge is labelled fail.
An End can’t have outgoing edges.

Checks

The Checks panel reports problems as you edit. Click a finding to select its step. Two kinds exist.

Problems that block saving. These are structural and show a red blocks saving tag, and Save stays off until they are fixed. Examples: a required field left empty, more than one outgoing edge on a step that isn’t a Decision, a fail edge on anything but a Verify, an edge out of an End, too many steps or edges. The server refuses a bad graph too, and the save-failed message points you to the Checks panel.

Findings. These never block saving. They are computed against the skill, its bundle and the chosen agent.

Severity	Finding
Error	A step not reachable from the start.
Error	A Decision with fewer than two outgoing edges, or an edge with no label.
Error	A Verify with no `fail` edge.
Error	A Tool the chosen agent doesn’t have.
Error	A File step whose path isn’t in the bundle.
Error	A Delegate target the chosen agent doesn’t have. For the `skill` kind this shows as a warning, because system skills aren’t checked.
Error	A compiled body over the size cap. This one also refuses the save.
Warning	A Step whose text ends in a question. Use Ask the user so the agent waits for the answer.
Warning	An Ask the user step when the agent runs on a channel that can’t show the question. The web app can.
Warning	No End reachable from a step.
Warning	A description under 40 characters.
Warning	A trigger phrase another skill in the project also uses.

With no agent chosen, the panel says the tool, channel and delegate checks are skipped.

Try it

Try it runs your unsaved draft on a real agent, so you can see how it behaves before you save or publish.

Pick an agent in the top bar. The skill must already be attached to that agent, and the agent must be running.
Type a Prompt. The skill name is added for you. Use Attach files to give it input files.
Click Run. It stays off while the graph has blocking problems or a run is in progress.

This is a real session on the agent. It uses the agent’s model like any chat. Nothing is published: the draft travels as an override for that one session. If the runtime doesn’t allow draft overrides, the panel says the published version ran.

While it runs, steps light up on the canvas, and a list shows each step visited, its time, and any tool calls. Stop watching stops following the run but not the session. Open session opens it on the Skill path tab.

Bundle files

Bundle files manages the supporting files a File step points at. Pick a folder (scripts/, templates/, references/ or assets/), click Add files, and upload. Each file in the list has Edit (text files only) and Delete. Changes apply to the skill’s repository at once, not when you click Save.

How the agent follows it

The compiled instructions are numbered steps, each headed with its id, such as Step 3 (s3). The agent is asked to report each step as it starts, finishes or skips it, with the skill_step tool.

The graph is not enforced. Nothing stops the agent from stepping outside the flowchart, and the canvas doesn’t offer skill_step as a tool to place in a step.

The Skill path tab

When a session used a procedure skill, its detail view gets a Skill path tab beside the thread. It shows the skill’s flowchart with the steps the agent took, in order, with the time on each. Beside it, a list gives each step’s number, type, time and tool calls (failed calls in red), then a count of steps skipped and steps never reached.

The tab shows the first procedure skill the session used. If the agent never reported its steps, it says “The agent did not report its steps, so the path cannot be shown.” A session that used no procedure skill has no tab.

Limits

At most 500 steps and 1000 edges per graph.
Skill optimization isn’t available for canvas skills.

To move a canvas skill between projects, use Export, then New skill and Import under Write markdown. See Export a skill.