Skip to main content
Use this pattern when clarity matters more than speed. By forcing the assistant to echo back its understanding, you surface misalignments early and avoid rework.

​
Why it works

  • Restatements expose missing details, incorrect assumptions, and shaky intent.
  • Shared alignment keeps implementation, testing, and validation on the same page.
  • Teams gain an auditable record of what was agreed before changes land.

​
How to prompt

  • End instructions with: “Iterate this back to me so I know we are aligned.”
  • Ask for a concise restatement of the goal, inputs, constraints, and expected outputs.
  • Require the assistant to flag uncertainties and propose clarifying questions before acting.

​
Example prompts

  • Building a feature: “Generate the flows for the new checkout. Summarize the intended screens, data dependencies, and edge cases, then iterate them back to me so I know we’re aligned.”
  • Fixing a bug: “Investigate the login timeout bug. Restate the bug, proposed fix, and the expected behavior change. Iterate back before touching code.”
  • Changing behavior: “We need a gentler rate-limit response. Describe current behavior and the proposed update, and iterate the plan back so we’re aligned.”

​
Quick templates

  • Alignment-first build: “Task: <what to do>. Scope: <files/areas>. Constraints: <perf/security/UX>. Output: <code/spec/tests>. Iterate back with: goal, plan, risks, tests. Ask clarifying questions before edits.”
  • Fragile area safety: “This area is critical. Before changes, restate current behavior, the intended change, files you’ll touch, and tests you’ll run. Pause for my confirmation.”

​
Bad vs. good (alignment)

  • Bad: “Fix the login bug.”
  • Good: “Login sometimes 500s after session refresh. Investigate auth refresh handlers. Restate the suspected cause, proposed fix, files to touch, and how behavior will change. Iterate back before edits.”

​
Guardrails for risky changes

  • Ask for a pre-change snapshot: current behavior, dependencies, and known edge cases.
  • Require an execution plan plus validation steps before edits.
  • Forbid unrelated changes: “Do not modify files outside <areas> unless confirmed.”