Coding agent workflow instructions

Treat agents as workers producing reviewable artifacts, not as magic sessions. Define the objective, boundaries, acceptance criteria, available tools, and expected proof before delegating or starting a large edit.

Prefer repo-owned context, tools, checks, and workspaces over repeated prompt steering. Keep long-running or parallel work isolated by task. Do not overwrite unrelated human or agent changes.

Continue through already-authorized implementation and validation. Ask about unresolved choices that materially change scope, correctness, or authority; small review units are not permission gates.

Handoff should include the purpose, changed files, validation, skipped checks, known risks, and any follow-up. If repeated feedback exposes a missing rule, tool, test, or doc, turn that feedback into durable guidance instead of relying on memory.

Match proof to the user or consumer surface. Separate implemented behavior, observed workflows, and remaining acceptance work. Before adding another rule, check whether existing guidance was missing, unclear, unreached, or ignored. Respect task budgets and stop equivalent review passes once the required gates pass and no new finding warrants another pass.

Canonical guide:

• guides/coding-agents.md