Claude-cron has 5 tools with complete Zod schemas and descriptions. Tool naming is verb-based and clear (schedule_task, list_tasks, cancel_task, run_now, install_daemon). All tools have non-trivial descriptions (194-500+ chars). However, there are critical issues: (1) payload parameter in schedule_task uses a generic object with optional nested fields rather than type-discriminated unions, making it ambiguous which fields apply to which task type; (2) parameter descriptions lack actionable constraints (e.g., 'when' accepts multiple formats but does not specify precedence or fallback behavior); (3) output schemas are not formally documented, responses are text-only Content blobs rather than structured objects with typed fields; (4) no error classification (retryable vs fatal) or recovery guidance in error paths; (5) install_daemon is marked IRREVERSIBLE but has no confirmation step or dry-run mode; (6) list_tasks returns formatted text rather than structured task objects, blocking downstream tool composition. Schema completeness is present but design is not aligned with Arcade's 54 patterns for response shaping, error recovery, and command safety.

• Replace the generic 'payload' object in schedule_task with type-discriminated unions. Define separate payload types for each task type (ReminderPayload, ShellPayload, PromptPayload, etc.) and use Zod's discriminatedUnion() to enforce field mapping based on the 'type' enum.
• Document output schemas formally. For schedule_task, return a structured object: {task_id, name, type, next_run, cron, status}. For list_tasks, return {tasks: [{id, name, type, status, next_run, last_run, run_count}], total, limit, offset}. This enables LLMs to chain tool calls without parsing text.
• Add error classification to all tool error responses. Categorize as 'retryable' (e.g., timeout, transient database lock), 'user_fixable' (e.g., invalid 'when' format, missing required field), or 'fatal' (e.g., permission denied, resource not found). Include a recovery suggestion: 'Try rephrasing the time as "2026-03-15 at 10:00 AM"' or 'Call list_tasks to get a valid task ID.'
• Add a dry_run parameter (boolean, default false) to install_daemon and a confirmation_token return value. Require the LLM to call install_daemon with dry_run=true first, review the summary, then call again with confirmation_token and dry_run=false. This prevents accidental system modifications.
• Clarify 'when' parameter parsing rules. Document: (1) parsing precedence (e.g., exact ISO 8601 match first, then natural language via chrono-node, then cron pattern); (2) timezone handling (default to system timezone or accept optional 'timezone' param); (3) invalid format handling (return specific error like 'Could not parse "invalid time". Try: "tomorrow 9:00", "2026-03-15T10:00:00", "0 8 * * 1"').