1 · Start from something that works
The quickest way in is a template. Open Sidebar → Plugins, scroll to Start from a template, pick one and give it a name. You get a plugin of your own that already runs:
Then change it. By hand in the plugin's folder, or by asking the chat: "Change the plugin My list so that every item has a due date." Asking about the same plugin again changes that plugin and keeps everything it saved.
A plugin can bring tools of its own, the way Snippets does: a small module in its tools folder, declared with a name, a sentence on what it does, the input it takes and what it does to the world (only reads, changes the plugin's data, or sends something outward). The chat can build one for you: "Give the plugin Snippets a tool that counts the words of a snippet." A tool runs inside what the plugin was granted and never wider, a tool that sends something asks you first, and a declaration with a bad name or a missing module is refused by name instead of half-written.
2 · What makes a plugin hold up
- A real result first. Make the thing the plugin should produce once by hand, and use it. That result is what the plugin is measured against.
- A gold standard. Pick one finished result that every check has to pass. If it fails a check, the check is wrong, not the result.
- Only the public interface. A plugin reaches Skales through its page interface and its own tools, nothing else. If something is missing, it is missing for everybody, and it is asked for, not worked around.
- Every view named before it is built. The full list of screens in order, and one sentence that says what good looks like. Failures say what went wrong, on the page.
- Real example content. What the page shows on its first open is real, not lorem ipsum, so a person sees at once what the plugin is for.
3 · The brief, before the first line
A brief is ready when every row has an answer. A row without one is a question to settle first, not a guess to build on.
| Part | What it says |
|---|---|
| Purpose | One sentence: who does what with it, and what comes out at the end. |
| Reference result | The real, hand-made result it is measured against, and who used it. |
| Gold standard | Which result every check must pass, and which checks those are. |
| Views | The complete list of screens, in order. |
| Quality bar | The rule in one sentence, plus the shared rule of the three modes below. |
| Page actions | Every action as a future tool of the plugin: its name, whether it reads, writes or sends, and where the approval sits. |
| Skales tools | The tools the plugin may call, each with a reason. An empty list means none. |
| Permissions | Network, files and memory. Keys and sign-ins go through Skales' encrypted store, never into the plugin's folder. |
| Example content | What appears on the first open. |
| Languages | English first, then the others the plugin carries. Product names stay as they are. |
| Example jobs | Two to five sentences a person could hand the AI in the "AI does it" mode. |
| Distribution | The public catalogue, or a file you hand on yourself. |
| Acceptance | One full run per mode, the gold-standard run, and the checks in section 6. |
4 · The three modes are a contract
Every plugin with a page carries all three. Skales draws the choice under the page, remembers it per plugin and tells the page before it paints.
| Mode | Who presses the buttons | What the page does |
|---|---|---|
| Manual | You | Marks every AI entry with data-skales-ai; Skales hides those in this mode. |
| With AI | You, the AI when asked | Shows its AI buttons; Skales adds a line to ask the plugin's agent. |
| AI does it | The AI | Offers its actions as tools of its own; every change is one step that can be undone. |
The rule all three share: the AI may do anything, you approve, and every step stays correctable on its own.
Where approval sits
- Send (sending, publishing, spending money): always an approval card, in every mode, even in the Unrestricted mode.
- Write to the page's own data: a proposal you take over, or an ordinary step in the page's history that Undo takes back. The page stays the only writer of its data; a tool proposes, it does not overwrite.
- Read: no approval.
5 · What every plugin page keeps
- Its data in the plugin's own storage or its own folder, never in browser storage.
- Every model call behind a button marked
data-skales-ai. - Every failure said on the page, in words a person can act on. With no model set up, the page says where to add one instead of showing an empty panel.
- The colours of the theme the person chose, from the variables Skales hands the page, and the app's own font.
- Its words in a language file per language when it carries more than English.
The page interface itself (storage, tools, model answers, jobs, previews, files) is described in the developer kit's plugin reference.
6 · Check it
- Ask the chat to test the plugin. Skales reads the manifest, checks that every page is whole, calls only what the interface answers, is allowed every tool it calls, keeps nothing in browser storage and marks its AI entries for the manual mode.
- Run it once in every mode, and once against your gold standard.
- Close the window in the middle of something and open it again: what you saw comes back. Do it twice in a row: nothing from the first run leaks into the second.
7 · Share it
- Export on the plugin's row on the shelf writes a
.skpluginfile: the definition and its files, never what the plugin saved while you used it. With a signing key on your machine the file is signed. - Hand the file to somebody, or publish it in the public catalogue at github.com/skalesapp/plugins, which lists where your release lives rather than copying it.
- Sharing from the export dialog can also post a line to Discover, if you take part in Discover.
A newsletter (lists, draft, design, test send, send, results), a small CRM (contacts, companies, pipeline, tasks, notes) or a shop helper (products, texts, pictures, prices, publish). In each, everything that sends or publishes asks you first.