Extensions
Six kinds of definition that teach Ginu your world, plus your vocabulary.
What this page covers
OutcomeCreate a small reusable skill, try it, then learn where vocabulary, connections and automation fit.
Before you beginPick a routine you already know well. Start with instructions that you can verify without a credential or an irreversible action.
Follow these steps
Extensions are definitions stored in your vault, in six kinds — skills, agents, integrations, surfaces, triggers and workflows — with Vocabulary on its own page beside them. Open one and read it before creating your own.
A skill is a reusable procedure: how you deploy, how this team writes a design note. It is a native Claude Code skill file, so Ginu manages it and the CLI is what loads and runs it. Describe one by voice, saying plainly when it should be used.
Open the saved skill and read its body, source and resource files. The use-when line is the part that matters most: it is what decides whether the skill is ever found, so make it specific enough to recognise.
Try it on one bounded task. If the behaviour misses your intent, correct the instruction by talking rather than rewriting it by hand, and try again before using it anywhere consequential.
An agent definition is a role rather than a procedure — instructions, tier, tools, and when it should be chosen — inlined into a worker’s prompt when one spawns with it. It is separate from the Agents screen, which shows the workers actually running.
An integration is a connection your agents can reach: an MCP server, or a command with a check and a fix. The definition declares credential slots and never values; paste a secret into its slot on the detail page, where it goes to the keychain.
Vocabulary biases the transcriber toward a word it mishears and gives the synthesiser a respelling for saying it well. Put the mis-transcription it actually produces in hear, then test the entry in one short phrase.
Example
Create a release-review skill. Use it when preparing a release: run the project’s tests, draft a concise note, and ask before publishing.
Open Skills and inspect the saved instruction and its use-when description. Then ask Ginu to use it on a small example and review the actual result.
A small loop you can repeat
- 01Describe a routine
- 02Inspect the skill
- 03Try a small task
- 04Refine the instruction
Start with the words Ginu should know
Vocabulary has its own page under Extensions. Use it for a product name, a person’s name or a term that is repeatedly misheard. The fresh vault’s Ginu entry is a worked example you can follow.
Written is the correct spelling. Spoken is a respelling that helps pronunciation. Hear contains the mistakes the recogniser actually makes. Test your entry in a short phrase, then check the transcript and spoken reply.
The budget strip names entries that do not fit in the current recognition prompt. Raise an important entry’s priority if needed. A common-word alias affects every matching occurrence, so choose one deliberately.
Give a routine a skill, and a role an agent
A skill is a reusable procedure written down once — how you deploy, how this team writes a design note. It is a native Claude Code skill file: Ginu stores and manages it, the CLI is what loads and runs it. Ask Ginu to create or edit one by conversation, then open its detail page to read the body, source and resources. The use-when line is what decides whether the skill is ever found, so make it specific enough to recognise.
An agent definition describes a role: its instructions, tools and when it should be chosen. Use New agent to create a definition, review the fields and save only when the scope is right. These definitions are separate from Agents, which shows the workers currently running.
Try the definition on one bounded task. If its behaviour misses the intent, refine the instruction and try again before using it across more work.
Connect a tool or create a useful view
An integration connects an MCP server or a command. Choose the kind first when creating one. Add secret values only in credential slots on the detail view; those values go to the keychain. Do not dictate a secret into the conversation.
Ask Ginu to probe or reconnect the integration, then test a small action. Deleting the definition does not delete its keychain entries.
A surface is a visual renderer for data, such as a project summary. Describe the view you want by voice and refine it in the designer while watching its sample-data preview. The source is read-only; you author the view through conversation.
Understand existing triggers and workflows
A trigger watches a schedule or event, checks an optional condition and takes its configured action. Open a row to inspect its source, raw definition and history. An error marker explains a failing check; read it before relying on the watch.
A workflow joins steps into a resumable job. Open an existing card to see its canvas. Select a step to edit its fields, add or remove steps, and inspect warnings and run history. Changes to fields save when you leave the field.
Creating workflows, triggers and hooks by voice is not connected yet. Triggers cannot launch workflows, and an already-busy workflow does not queue another run. Build your routine around the capabilities available in the installed build.
Move and maintain a set of extensions
Use the gear on an Extensions tab to open pack settings. Install a pack from a file, pasted text or a URL. If an item already exists, choose whether to replace it, keep yours or install alongside it.
Export lets you choose items by kind. Credential-shaped values are left out, so plan to configure the destination’s credential slots separately.
Review update notices and differences before replacing your own changes. Automatic updates are off by default. Update checking and schema migration are separate settings: one finds new source content; the other updates the format of your existing items.
Notes
- A surface is a small named renderer — an HTML template Ginu fills with data and shows in a sandboxed frame. It is authored by conversation; there is no editor for one anywhere in the app.
- A trigger watches a schedule or an event, runs an optional cheap check and files a task. A workflow strings steps into one resumable job. Both can be read and edited here; creating either by voice is not connected yet, and a trigger cannot launch a workflow.
- Do not dictate a credential. Spoken text can become part of a transcript.
- The character count under the cards is what Extensions costs in every prompt, and a capped list means some installed items are no longer named to the model. Automatic pack updates are off by default.
Check
You have inspected one definition and tried it on a small task. You can explain when it should be used and what still requires your decision.