Prompts
Saved prompt templates you can drop into any conversation.
Prompts are saved query bodies for the Mel agent — the instructions you find yourself retyping ("review this diff for security issues and…", "summarize the failing tests and propose fixes…") kept as reusable objects in Mel Drive. They're stored as *.prompt.json files under ~/.local/share/mel/drive/.
Creating a prompt
- Open the tools panel (
ctrl+\), switch to the Drive view, and click + → Prompt. - The prompt editor opens in the center panel with three fields:
- Title — how it appears in the Drive tree
- Description — a one-liner so future-you knows what it's for
- Body — the actual prompt text
- Click Create.
The same editor opens when you click into an existing prompt, so refining a prompt over time is the same motion as writing it.
Using a prompt
Click a saved prompt in the Drive tree and Mel inserts its body as the query in the active pane. From there it behaves like anything you typed yourself — edit it, add specifics for today's situation, and submit. The prompt object is a starting point, not a fire-and-forget macro, which is usually what you want: real prompts almost always need one situational detail appended.
Hovering a prompt in the tree shows the standard Drive preview card — title, a snippet of the body, and when it was last edited — so you can tell similar prompts apart without opening them.
What makes a good saved prompt
- Encode the standards, not the specifics. "Review for error handling, injection risks, and missing tests; report as a checklist" ages well. "Fix the bug in auth.rs" doesn't.
- State the output shape. Prompts that specify the format you want back (a table, a checklist, a diff) are the ones worth saving.
- Keep the variable part for typing time. If a prompt needs the same slot filled every run, leave an obvious gap at the end and fill it after insertion.
Prompts vs workflows
Both insert text into your prompt bar, but they target different executors:
- A workflow is a shell command with
{argument}placeholders — it inserts with the first argument selected because commands have structured slots. - A prompt is natural language for the agent — it inserts as-is for you to extend.
If you're saving ffmpeg flags, you want a workflow. If you're saving "how I like code reviewed", you want a prompt.
Related
- Mel Drive overview
- Notebooks — for longer-form knowledge that isn't a single query