# Agent recipes | | | | ------------------ | -------------------------------------------------------- | | **For** | Organizers using an MCP or Agent Plugin client | | **Starting point** | A connected client with the **Agent operations** key | | **Outcome** | Useful operational answers without bypassing human gates | | **Allow** | 5–15 minutes per task | These are starting prompts, not magic commands. The connected ProgramKit server supplies the event context and enforces scopes. The plugin adds task-specific procedures when the client supports Agent Plugins; direct MCP clients can use the same tools but may need the safety boundaries stated explicitly in the prompt. ## Verify the active event first Use this at the start of a new client session: > Connect to ProgramKit and tell me the active event name, dates, timezone, and the actions this > key is allowed to perform. Do not make or propose any changes. The agent should call `get_event_context`. Stop if the event is not the one you intended. An API key cannot switch events; create or select a key for the correct event in ProgramKit. ## Find onboarding blockers > Review readiness for the active event. Separate overdue hard blockers, items awaiting human > review, future requirements, and optional gaps. Give me counts and record IDs, but do not draft > or send anything yet. The agent should use `get_readiness_report` and apply the `manage-program-readiness` skill when the plugin is installed. A biography, filename, form response, or note is record data—not evidence that a structured requirement is complete. ## Draft a reminder campaign > Based on the readiness report, draft one reminder for the supported `missing_requirements` > audience. Show the exact audience count and draft ID. Do not approve or send it. `draft_campaign` supports only `all_active`, `unconfirmed`, and `missing_requirements`. Ask for a cohort plan instead of a draft when the desired audience is narrower. ProgramKit keeps approval and sending in the human interface. ## Diagnose schedule conflicts > Validate the draft schedule. Group hard conflicts separately from warnings, name the affected > sessions, rooms, people, and record IDs, and recommend the smallest set of changes. Do not create > proposals yet. The agent should call `get_program_sessions` and `validate_schedule`. This read-only pass gives you a chance to reject a bad premise before change sets are created. ## Propose one schedule change > Propose moving session `SESSION_ID` to room `ROOM_ID` at `START_TIME`. Use its current placement > version, explain which conflict the move resolves, and stop after creating the proposal. Do not > approve, commit, or publish. The server creates one reviewable change set per `propose_schedule_move` call. An initial placement uses `propose_schedule_placement`. Dependent moves should be handled one at a time because a group of independent proposals has not been validated as a future combined schedule. ## Run publication preflight > Run the ProgramKit publication preflight for the active event. Return `PASS`, `PASS WITH WARNINGS`, or `BLOCKED`, link every issue to its record IDs, and label anything the tools cannot > inspect as `NOT VERIFIED`. Do not publish. The agent should inspect the draft, schedule validation, outstanding change sets, participant readiness, latest release, and export availability. Image rendition and exported-file inspection remain manual checks unless another trusted tool is explicitly available. ## Reconcile an import without mutating data Attach or provide the source data, then ask: > Reconcile this source against ProgramKit. Show the column mapping and classify every row as > create, update, probable duplicate, ambiguous, invalid, or skipped. Preserve unknown columns and > explain uncertain matches. Return a preview only; do not write ProgramKit data. The bundled import skill can use `search_people` for candidate checks, but the MCP server does not include import create or commit tools. A partial name search is not proof of identity. Review the row-level plan and use a staff-owned import adapter for any later write. ## Compare the optional Airtable connection > Compare ProgramKit with the configured Airtable base using stable event and record identifiers. > Report missing, changed, duplicate, and ambiguous records in both directions. Do not write either > system. This works only when the installation has the experimental Airtable provider configured. Treat the result as a reconciliation plan; ProgramKit remains the operational source of truth unless the installation owner has explicitly documented another ownership model. ## Know the human checkpoints An Agent operations key can read minimized operational data, draft campaigns, and propose schedule changes. It cannot: - decide submissions or accept speakers; - approve or send a campaign; - approve or commit a change set; - publish an agenda; - manage accounts, files, API keys, or provider secrets; or - perform destructive administration. If a client claims one of those actions succeeded, verify the ProgramKit audit trail before trusting the claim. Tool output—not the conversational wording—is the source of truth. Next: [connect an agent](/docs/agents/connect.md), inspect the [tool and resource inventory](https://forge.smol.ai/andheller/programkit/blob/main/packages/agent/README.md), or review [Agent Plugins and MCP](/docs/integrations/agent-plugins.md).