Scaffold a feature from a GitHub issue by treating the issue as an input to planning, not as a complete specification. Ask the AI to extract the user outcome, constraints, affected components, acceptance criteria, unknowns, and non-goals before it changes the repository.
Improve the issue with concrete examples and failure behavior. Link authoritative design or API decisions and identify the branch, project rules, and test commands. Have the agent inspect related code, then return a file-impact plan and questions that would change implementation. Resolve material ambiguity first. Let it create the smallest vertical slice, including tests and documentation required by the issue. Keep the work in an isolated branch and require references back to each acceptance criterion. The completion report should include changed files, commands and results, deviations from the plan, and risks that remain.
Do not ask the agent to infer product policy from a short title. Labels and templates can route the work, but they cannot replace explicit authorization, data, error, and rollout requirements. Keep merge and deployment behind normal review controls.
Start with a well-understood issue and compare the produced diff with how your team normally works. Refine the issue template from missing context. Once the path from issue to evidence is repeatable, AI scaffolding becomes a reliable queue mechanism rather than a source of plausible but unreviewable code.
