A repository contains code, abandoned ideas, generated output, and instructions that may disagree. Handing over more text does not resolve those disagreements. Give the assistant a route through the evidence and tell it what counts as authoritative.
Start with the boundary
Name the failing behavior, the entry point, and the constraint that matters most. Include the smallest reproduction and its exact output. Separate what you observed from what you suspect so your theory does not become an accidental requirement.
Assemble a context packet
- Goal: the observable behavior that should change.
- Evidence: relevant files, a failing check, and the input that triggers it.
- Constraints: public interfaces, supported environments, and scope limits.
- Unknowns: questions the assistant should investigate before editing.
Leave a trail, not a dump
Point to a nearby implementation that follows the repository conventions. Summarize long logs, retaining the relevant error and enough surrounding lines to interpret it. Remove credentials and irrelevant personal data before sending material to any endpoint.
Refresh when the facts change
After an edit, replace stale observations with the current diff and check output. Ask the assistant to identify remaining uncertainty. When a conversation becomes tangled, a short handoff with the current state is often easier to inspect than a replay of every attempt.
