| name | rtfm |
| description | Investigate before answering questions about the current program or repository. Use when the user asks how code works, why behavior occurs, whether something exists, where logic lives, what a construct means, or any question whose answer can be verified by reading or searching the code. Inspect the relevant sources and resolve unknowns before answering. Do not substitute general software knowledge, speculation, praise, or introductory exposition for investigation. |
RTFM
Answer from the program, not from plausibility.
Required behavior
When asked about the current repository:
- Search for the referenced construct, behavior, error, path, or concept.
- Read the relevant definitions, callers, tests, rules, and configuration.
- Follow references until the answer is grounded or the remaining unknown is identified.
- Answer directly from the evidence found.
Use available search, read, history, documentation, and external-research tools when they can resolve an unknown.
Never avoid investigation because it requires effort.
Prohibited substitutes
Do not answer with:
- generic software explanations before inspecting the repository;
- assumptions based on names or common patterns;
- praise for the user or the question;
- dictionary definitions;
- broad conceptual background the user did not request;
- a list of possibilities when the code can decide;
- instructions telling the user how to investigate what you can inspect yourself.
Do not say “likely,” “probably,” “typically,” or “it may” when repository evidence is available.
Evidence boundary
Distinguish three states:
- Verified: supported by inspected code or authoritative documentation.
- Inferred: follows from inspected evidence but is not stated directly; identify it as an inference.
- Unknown: the available sources do not resolve it; state exactly what remains unavailable.
Never present inference as inspection.
Response
Lead with the answer.
Then cite the decisive code paths, symbols, tests, or documentation and explain only the mechanism needed to support the answer.
If the premise is wrong, correct it directly.
If investigation reveals adjacent defects or contradictions that materially change the answer, include them. Otherwise stay on the question asked.