The prompt board
The prompt board is a kanban for AI prompts. Open it by clicking Prompts in the global rail. It has two tabs: Board (kanban) and Library (templates).
The kanban board
Cards move through four columns:
| Column | Meaning |
|---|---|
| Idea | Quick captures — a prompt you want to run eventually |
| To do | Written and ready to run |
| In progress | Currently running or being iterated |
| Done | Completed runs |
Drag a card between columns to change its stage. Running a card moves it to In progress automatically, then to Done on success.
Creating a prompt card
Quick capture: click + at the top of the Idea column, type a title, press Enter.
Full card: click any card to open the card editor, where you set:
| Field | Description |
|---|---|
| Title | Short label shown on the board |
| Kind | feature · fix · refactor · test · explain · custom |
| Instruction | The prompt text — what you want the model to do |
| Template | Start from a saved template to pre-fill instruction and kind |
| Connected docs | Workspace docs to include as context (with token counts) |
| Target | Where the prompt runs |
Variables
Use {{variable}} placeholders in the instruction text. When you run a card with
variables, Bedrock asks you to fill them in before sending.
Example instruction:
Refactor the {{component}} component in {{file}} to use the new {{pattern}} pattern.
Follow the rules in CLAUDE.md.
When you run this card, Bedrock prompts for component, file, and pattern.
Connected docs
Tick workspace docs to include as context. The combined token count is shown — same engine as the Context Composer. This makes prompts self-contained: the right context is always attached, even weeks later.
Targets
A prompt’s target decides what Run does:
| Target | What happens |
|---|---|
| Clipboard | Copies the assembled payload to paste anywhere |
| Claude (Terminal) | Opens Terminal, launches Claude interactively (macOS only) |
| Claude CLI | Runs Claude headlessly; captures the reply in the card. Can edit files. |
| Claude API | Calls the Anthropic API — text reply only |
| Gemini API | Calls the Gemini API — text reply only |
| LM Studio | Runs a local model — fully offline, no key |
Only targets that are enabled in Settings appear in the list. See AI integrations for setup instructions.
Running prompts
Single card: open it and click Run. Fill in any variables. The reply appears in the card when done.
Whole column: click Run all at the top of To do or In progress to run every card in that column in sequence.
The ✨ Generate menu
In the card editor, ✨ Generate opens the writing-assist menu:
| Action | What it does |
|---|---|
| Write from title | Drafts a full instruction from the card’s title |
| Improve | Makes the instruction clearer and more effective |
| Rephrase | Rewrites with different wording |
| Shorten | Trims to the essential |
| Custom… | Apply any instruction of your own |
Requires LM Studio — see Writing assist.
The Library tab
Shows all available prompt templates: built-in templates, the bundled Claude Code pack, and your own saved templates. Click a template to apply it — creates a new prompt card in To do with the template’s instruction and kind pre-filled.
Tips
- Keep instructions short and concrete. Vague prompts produce vague results.
- Always attach at least your
CLAUDE.mdas a connected doc so the model knows your conventions. - Use the Claude CLI target for tasks involving file edits — it’s the only target that can write to your project.
- Save recurring prompts as templates so they’re available across workspaces.