Task solving
Read the question and any supplied context carefully. Identify the requested entity or value, check the supporting evidence, and answer concisely. Follow any explicit output-format requirements. Describe tool use only when it actually occurred, and distinguish supplied evidence from assumptions.
AppWorld tasks
When executing AppWorld tasks through mcp__appworld__appworld_execute, use the preloaded API client and manage the turn budget.
API client and discovery
- Use the preloaded
apisglobal directly; do notimport apis. If an import error occurs, the client is already available and can be inspected withprint(dir())orprint(sorted(globals().keys())). - Use preloaded helpers such as
json,Counter,defaultdict,datetime, andrequesterinstead of restricted imports or filesystem/introspection helpers such asos,sys,inspect, oropen. Avoid blocked or unnecessary imports/introspection; they fail or waste turns. - Inspect only the needed API docs before acting:
apis.api_docs.show_app_descriptions(),apis.api_docs.show_api_descriptions(app_name=...), andapis.api_docs.show_api_doc(app_name=..., api_name=...). Print the results and note app names, parameter names, required fields, response shapes, and documented limits. - In time-sensitive tasks, obtain the simulated current date/time from world APIs or runtime time globals rather than assuming from token expiry, file timestamps, or general knowledge.
Authentication and data access
- Get supervisor context early:
apis.supervisor.show_active_task(),apis.supervisor.show_profile(), andapis.supervisor.show_account_passwords(). - For authenticated app APIs, inspect the target app's login API doc, log in once with the supervisor-provided app account email/username and password, save the returned
access_token, and reuse it for all subsequent authenticated calls to that app. Missing tokens commonly produce 401 errors. - Before creating records, use list/show APIs to check for existing resources and avoid duplicates. For create/update, preflight reads: list/show all relevant records across pages first.
- Use detail/show endpoints when list responses omit fields you need. Inspect raw records before accessing nested, numeric, or optional fields.
- Resolve people/accounts through available search endpoints and identify current records by flags such as
is_currentor position fields; when a task refers to "so far" in an ordered or cyclic structure, operate on the prefix ending at the current item. - For file/email evidence, list relevant directories/mailboxes and read readable records; ignore unreadable binary duplicates, watch for overlapping file formats, and cross-check current-world dates from messages, app APIs, or task context before applying relative terms.
- For list-returning endpoints, page through all results using the documented pagination parameters (usually
page_indexandpage_limit). Respect documented caps such aspage_limit <= 20, and continue until an empty or short page is returned; collect all pages before filtering, sorting, or checking existence. Do not assume page 0 contains all records, that one page is complete, or that results are sorted.
Execution and completion
- Batch independent actions inside as few
mcp__appworld__appworld_executecalls as practical: fetch credentials and token once, then loop over records in a single Python block. Use server-side filters where available, but still verify after listing or mutating state. - Once enough evidence is available, stop exploring and execute the required actions; avoid extra exploratory calls that may exhaust the turn budget. If near the turn limit, submit the best supported answer instead of over-exploring.
- After state changes, verify the final state with read/list APIs.
- Read the supervisor completion API doc and follow its answer contract exactly. For question tasks, pass only the concise requested value as
answer; avoid adding units, labels, or prose. For action-only tasks, call without an answer or passNoneand do not summarize actions. Use the documented status value (usually"success") and callapis.supervisor.complete_task(...)promptly. - Use only the supplied AppWorld APIs; do not access grading, ground-truth solutions, or benchmark files.
name: main description: General guidance for solving tasks and following requested output formats.
Task solving
Read the question and any supplied context carefully. Identify the requested entity or value, check the supporting evidence, and answer concisely. Follow any explicit output-format requirements. Describe tool use only when it actually occurred, and distinguish supplied evidence from assumptions.
AppWorld tasks
At the start of an AppWorld task, read this skill with the Skill tool before calling mcp__appworld__appworld_execute. Use the preloaded API client and manage the turn budget.
API client and discovery
- Use the preloaded
apisglobal directly; do notimport apis. If an import error occurs, the client is already available and can be inspected withprint(dir())orprint(sorted(globals().keys())). - Use preloaded helpers such as
json,Counter,defaultdict,datetime, andrequesterinstead of restricted imports or filesystem/introspection helpers such asos,sys,inspect, oropen. Avoid blocked or unnecessary imports/introspection; they fail or waste turns. - Inspect only the needed API docs before acting:
apis.api_docs.show_app_descriptions(),apis.api_docs.show_api_descriptions(app_name=...), andapis.api_docs.show_api_doc(app_name=..., api_name=...). Print the results and note app names, parameter names, required fields, response shapes, and documented limits. - In time-sensitive tasks, obtain the simulated current date/time from world APIs or runtime time globals rather than assuming from token expiry, file timestamps, or general knowledge.
Authentication and data access
- Get supervisor context early:
apis.supervisor.show_active_task(),apis.supervisor.show_profile(), andapis.supervisor.show_account_passwords(). - For authenticated app APIs, inspect the target app's login API doc, log in once with the supervisor-provided app account email/username and password, save the returned
access_token, and reuse it for all subsequent authenticated calls to that app. Missing tokens commonly produce 401 errors. - Before creating records, use list/show APIs to check for existing resources and avoid duplicates. For create/update, preflight reads: list/show all relevant records across pages first.
- Use detail/show endpoints when list responses omit fields you need. Inspect raw records before accessing nested, numeric, or optional fields.
- Resolve people/accounts through available search endpoints and identify current records by flags such as
is_currentor position fields; when a task refers to "so far" in an ordered or cyclic structure, operate on the prefix ending at the current item. - For file/email evidence, list relevant directories/mailboxes and read readable records; ignore unreadable binary duplicates. Do not repeatedly attempt to read files returned as
binary:...; instead use readable text/alternative records or proceed with available evidence. - For list-returning endpoints, page through all results using the documented pagination parameters (usually
page_indexandpage_limit). Respect documented caps such aspage_limit <= 20, and continue until an empty or short page is returned; collect all pages before filtering, sorting, or checking existence. Do not assume page 0 contains all records, that one page is complete, or that results are sorted.
Execution and completion
- Batch independent actions inside as few
mcp__appworld__appworld_executecalls as practical: fetch credentials and token once, then loop over records in a single Python block. Use server-side filters where available, but still verify after listing or mutating state. - Once enough evidence is available, stop exploring and execute the required actions; avoid extra exploratory calls that may exhaust the turn budget. If near the turn limit, submit the best supported answer instead of over-exploring.
- After state changes, verify the final state with read/list APIs.
- Read the supervisor completion API doc and follow its answer contract exactly. For question tasks, pass only the concise requested value as
answer; avoid adding units, labels, or prose. For action-only tasks, call without an answer or passNoneand do not summarize actions. Use the documented status value (usually"success") and callapis.supervisor.complete_task(...)promptly. - Use only the supplied AppWorld APIs; do not access grading, ground-truth solutions, or benchmark files.