Server has 8 well-named tools with complete input schemas and descriptions. All tools follow verb_noun naming (slack_list_*, slack_post_*, slack_get_*). Descriptions are present and actionable (avg ~80 chars). However, output schemas are completely undocumented, no tool declares what fields it returns, forcing LLMs to guess. Parameter descriptions are adequate but lack constraints (enums, ranges, formats). No error handling guidance, no recovery paths, no pagination documentation despite two tools supporting cursors. Missing tool annotations (readOnlyHint/destructiveHint) despite clear risk classification. No documentation of which tools are idempotent or safe to retry.

• Add output schemas to all 8 tools. Document return type (object), required fields (e.g., channel_id, message_ts, user_id), and field types. Example: slack_post_message returns {ok: boolean, channel: string, ts: string, message: {type: string, user: string, text: string}}.
• Add tool annotations: set destructiveHint=true on slack_post_message, slack_reply_to_thread, slack_add_reaction; set readOnlyHint=true on all get_* and list_* tools. This enables agents to reason about side effects.
• Add error recovery hints to descriptions. E.g., slack_post_message: 'If channel_id is invalid, call slack_list_channels() to find valid channel IDs.' slack_get_user_profile: 'If user_id is not found, call slack_get_users() to search by name.'
• Document pagination behavior. For slack_list_channels and slack_get_users, clarify: 'cursor is opaque; pass the next_cursor from the previous response. Omit cursor on first call. Stop when next_cursor is null or absent.'
• Replace verbose parameter descriptions with formal constraints. E.g., thread_ts: 'Slack message timestamp in format NNNNNNNNNN.NNNNNN (10 digits, period, 6 digits). Example: 1234567890.123456.' reaction: 'Emoji name without colons (e.g., thumbsup, heart, laughing). See https://www.webfx.com/tools/emoji-cheat-sheet/ for valid names.'