Finance-assistant has 12 tools with complete input schemas and descriptions. Naming follows verb_noun convention (import_preview, get_totals, commit_import, etc.). Descriptions are substantive (100-250 chars typical), explaining what each tool does and when to use it. However, output schemas are not documented, LLMs cannot see what fields to expect from responses. Error handling is minimal: tools return generic dicts with 'error' keys but lack recovery guidance or categorization. Parameter descriptions are present but sometimes lack constraints (e.g., account_id accepts any string; no enum of valid accounts). The commit_import tool implements a dry-run-first pattern (requires expected_to_import count), which is excellent for safety. Overall: solid foundation, but missing output documentation and error guidance that would push this to 75+.

• Document output schemas for all tools. Add a 'Returns' section to each tool docstring describing the response structure, e.g., 'Returns {income: {category: amount}, expenses: {category: amount}, total_income: number, total_expenses: number}'.
• Add error categorization and recovery hints. When returning errors, include a 'recovery' field: {error: 'Invalid currency', recovery: 'Valid currencies: EUR, USD, GBP. Call get_accounts() to see configured currencies.'}
• Convert free-form string parameters to enums where valid values are known. For account_id, currency, locale, and step, define enums in the schema or add a discovery tool (e.g., list_accounts, list_currencies) that agents can call first.
• Implement confirmation for destructive writes. For add_account and add_transaction, add a dry_run parameter (default true) so agents preview changes before committing. Or require a confirmation_token from a separate confirm_* tool.
• Add per-item error reporting for batch-like operations. If onboard_step processes multiple fields, return {parsed: {...}, errors: [{field, reason}]} so agents know which parts succeeded and which need retry.