Prompts

Prompts are reusable, named workflows the server ships to the client. Pick one (or the agent will), fill in the arguments, and the agent runs the steps using the right tools and resources.

In Claude Desktop, Claude Code, Cursor, Windsurf, and Zed they surface as slash commands: /triage-spam, /summarize-week, …. Or just ask the agent: "Run the triage-spam prompt for Contact."

Catalogue

Prompt	Arguments	What it does
`triage-spam`	`form_id`, `since?`	Walk spam folder, decide keep/restore/delete, apply on confirm
`summarize-week`	`form_id`	Last 7 days: themes + counts, markdown summary
`draft-autoresponder`	`form_id`, `tone?`	Drafts a body, previews, saves only on confirm
`audit-webhooks`	`form_id?`	Health-check every webhook, bucket failures, offer to replay
`categorize-submissions`	`form_id`, `since?`	Backfill `category` on uncategorized submissions
`draft-form`	`purpose`, `audience?`	Sketch a new form schema (returns JSON; doesn't ship)
`weekly-digest-brief`	`form_id`	Email-ready brief: headline, spam delta, top themes, outlier
`integration-health`	none	Per-team audit of webhooks + autoresponders

Why prompts (vs. ad-hoc asks)

A prompt is a guardrail. The agent does what the prompt says, not "whatever it thinks helpful." Use prompts for repeated, consequential workflows (spam triage, webhook audits). Use ad-hoc prompts (see Examples →) for one-off questions.

Two patterns the prompts share:

Plan first, apply on confirm. Anything that mutates state shows you the plan and waits for your approval.
Drafts don't ship. draft-form and draft-autoresponder produce previews; you decide when to call create_form / update_form.

What's next

Tools → — the action surface
Resources → — read-only data prompts pull from
Examples → — paste-ready ad-hoc prompts