Prompts and skills
The Prompts button opens one panel with three tabs. It is the same panel on a terminal's composer and on the orchestrator input in the Terminals view.
The panel
Prompts sits at the right end of the composer. It shows a count while skills are staged on the next prompt. On a terminal the panel spans the terminal's width and opens below it when there is room, otherwise above it. It keeps one height across its tabs, so switching tabs never moves the rows you are reading. There is no search field.
| Tab | What it lists | When it is empty |
|---|---|---|
| Recently used | Prompts you already sent, newest first | nothing yet |
| Saved | Prompts kept as templates | no saved prompts |
| Skills | Every skill on disk, all providers together | no skills found |
Recently used
Prompts you sent from your terminals and from the Terminals view input, newest first. The same text sent twice shows once, and the list holds the latest 40. Each row says when it was sent, such as 20m ago or yesterday, and which project it was sent in. A prompt from the Terminals view input shows no project.
Clicking a row puts its text into your draft, ready to edit and send again. Nothing is sent for you. Each row also carries the keep control described below.
Turns the app sent by itself, such as an auto continue, are not prompts you typed and never list here.
Keeping a prompt
A prompt you already sent can become a saved prompt. The keep control, a bookmark, sits on every sent prompt in a terminal's transcript, in the Terminals view conversation and in the WorkMap's conversations, and on every row of Recently used.
Keeping sends the prompt to the fallback model, which turns it into a template. The parts specific to that run become slots, and the general parts stay as they are. That takes a moment, so the control shows a working state and then a check. If it fails, the prompt reads could not save that one. A prompt longer than 8,000 characters cannot be kept. Keeping a prompt under a name that already exists replaces that saved prompt rather than adding a copy beside it.
Saved prompts
The Saved tab lists your saved prompts, the most recently used first, then the most recently made. A row shows the name, the description, and a line reading Model to be used with the model and effort it was written for. Beside it are an edit control and a remove control.
- Clicking a row stages the prompt, and the row shows a tick. It rides your next run as a file and shows in the dock under its own glyph. Click the row again to take it off. Staging a prompt counts as using it, so it moves to the top.
- Edit opens the file in the attachment window. Click the text to edit it. It saves back to the file on your next click outside it, or on Escape. The description, the model and the effort are part of that text, so you change them there too, and the row reads the file again once the save lands.
- Remove deletes the saved prompt at once.
Nothing is filled in for you. Type the details of this run into the input, and the model reads your request and the template together.
Each saved prompt is one .prompt.md file in the prompts folder inside the data folder. A phone paired through the mobile remote lists the same saved prompts.
The model a saved prompt names
Every saved prompt names a model and a reasoning level. It is the model the original request asked for when it asked for one. Otherwise it is the default model of the provider that ran it. The level follows the same rule, and whatever is picked lands on a level that model actually offers. Editing the file changes both.
Skills
The Skills tab lists every skill on disk, from every provider, in one list. Each row carries the mark of the provider whose format it is. Skills from Warshall's own skills folder carry no mark. A global tag means the skill comes from outside the project, such as your home folder or an account's folder.
Clicking a skill stages it, and the row shows a tick. It rides the next prompt for the provider to load by its own rules. Click it again to take it off.
On a terminal the list is that project's skills. On the orchestrator input it is every open project's skills together.
Where skills come from
| Provider | Looked for in |
|---|---|
| Claude | .claude/skills in the project, the same in your home folder, each linked Claude account's folder, and every skills folder inside your Claude plugin marketplaces |
| Codex | .codex/skills and .codex/prompts in the project, the same two in your home folder, and both in each linked Codex account's home |
| Cursor | .cursor/skills in the project, the same in your home folder, and .cursor/skills-cursor, the set Cursor installs with itself |
| Shared | The skills folder inside Warshall's data folder |
A skill is a folder holding a SKILL.md. Codex and the shared folder also read the older shape of one markdown file per prompt. OpenCode skills are not listed yet.
Names are deduplicated per provider. A project skill hides one of the same name in your home folder, which hides one from an account.
Skills compiled into a tool's own binary cannot be listed. Claude Code ships its built in set that way, so the model can use them and this panel cannot show them. Everything you or a plugin put on disk is listed.