Ask a question
1
State the result you need
Ask about the data as you would ask a person who knows the folder.You do not need to name a file, write SQL, or choose a tool.
2
Add constraints
Include the period, exclusions, grouping, or definition that matters. For example:
Exclude transfers and compare monthly spending in 2025 with 2024.3
Inspect and refine
Open Evidence to confirm the source and computation. If the interpretation is wrong, state the correction in a follow-up.
Match the question to the data
SQL is preferred because its result is compact and can often be re-run during verification. Python is an analysis escape hatch, not a file-management shell, and requires Python 3 on the computer.
Ask for the shape, not just the headline
For a simple lookup, ask directly:Follow-up questions
Follow-ups can reuse a bounded digest of the last three useful turns in that conversation, including recent headlines and successful SQL. Narrow a question when the first answer used the wrong interpretation:When Fella refuses
Past files cannot establish a future value. Questions such as “How many books will I finish next year?” should be declined rather than turned into a forecast. If the folder does not contain enough information, Fella says so. Definitions and general explanations can be answered without a tool. A leadingBackground: line marks general knowledge, not workspace evidence, and the prompt bars it from supplying specific figures.
Stop or retry
PressEsc or click the stop button beside the composer to stop. The run keeps evidence already collected and returns Stopped. While an answer is running, submitting a plain-text correction cancels it and re-asks with that correction. Use /retry after a transient provider error or after changing models.
Inspect the evidence
See the exact computation behind a response.
Learn the harness
Read the loop and its limits in source-backed detail.