GPT 6 FIELDGUIDE
GPT 6 PROMPT

Review a code change

Review a diff for correctness, regressions, and missing verification.

Independent resource

Complete prompt

Task
Review a diff for correctness, regressions, and missing verification.

Instructions
Review this diff for [REPOSITORY CONTEXT]: [DIFF]. Prioritize user impact, correctness, security, data loss, compatibility, and test gaps. Cite file and line. Do not praise style unless it affects behavior. If behavior cannot be established from the diff, state what must be checked.

Expected output
Findings ordered by severity, questions, and a focused verification plan.

Checks
- Findings are actionable.
- Severity reflects impact and likelihood.
- Unknown behavior is not asserted.

Expected output and checks

Expected output

Findings ordered by severity, questions, and a focused verification plan.

Checks

  • Findings are actionable.
  • Severity reflects impact and likelihood.
  • Unknown behavior is not asserted.

Worked example

Review a code change: recorded example

Codex · GPT-6 Astra · medium reasoning ·

Original input

Task
Review a diff for correctness, regressions, and missing verification.

Instructions
Review this diff for Small fictional ticket-pricing utility; cents integer,discount percentage0–100,input arrays must remain unchanged.: Synthetic TypeScript repository. Prices are whole cents, percent is 0–100, and callers retain ownership of the original array.
File src/tickets.ts, NEW line numbers shown:
1 export function discountedCents(cents: number, percent: number) {
2   return Math.round(cents * (1 - percent));
3 }
4 export function cheapestFirst(items: { cents: number }[]) {
5   return items.sort((a, b) => a.cents - b.cents);
6 }
Both functions are newly added. Review correctness, not formatting.. Prioritize user impact, correctness, security, data loss, compatibility, and test gaps. Cite file and line. Do not praise style unless it affects behavior. If behavior cannot be established from the diff, state what must be checked.

Expected output
Findings ordered by severity, questions, and a focused verification plan.

Checks
- Findings are actionable.
- Severity reflects impact and likelihood.
- Unknown behavior is not asserted.

What the checks found

  • The two seeded defects are correctly located at lines 2 and 5: percent needs division by 100; sort mutates the caller array.
  • Independent executable tests reproduce both failures and pass after the proposed changes. Other concerns are separated as questions, not reported defects.

Revisions after the first output

  • No correction to the review. A separate test receipt contains before/after values and assertions.
Original output

  1. [P1] Divide the percentage by 100 before applying it — src/tickets.ts:2. The contract supplies a value from 0 to 100, but the expression treats it as a fraction. For 1,000 cents at 20%, it returns −19,000 cents instead of 800; at 1%, it returns zero. This breaks ordinary pricing inputs and can produce negative prices. Change the expression to Math.round(cents * (1 - percent / 100)), subject to confirming the rounding rule below.

  2. [P1] Sort a copy to preserve the caller's array — src/tickets.ts:5. Array.prototype.sort reorders items in place and returns that same array. Calling this with [200, 100] changes the caller's sequence to [100, 200], directly violating the explicit ownership contract and potentially changing other displays or processing that share it. Use [...items].sort((a, b) => a.cents - b.cents). A shallow array copy is enough to prevent this sort from reordering the original; it does not promise independent item objects.

Questions to resolve from the surrounding contract, not assumed defects:

  • What rounding rule applies to fractional-cent discount results? Whole-cent output does not specify nearest versus floor/ceiling or tie handling.
  • Are nonfinite numbers, negative prices, fractional percentages and values outside 0–100 rejected by callers or by these functions? The diff does not establish validation responsibility.
  • What maximum cents value is supported? JavaScript number precision may matter at large values, but no such input range is supplied.
  • Are returned objects intended to be shared with the input, or does the API promise a deep copy? Only preservation of the original array is stated.

Focused verification plan:

  • Discount boundaries: discountedCents(1000, 0) === 1000, (1000, 100) === 0, (1000, 20) === 800, (1000, 1) === 990, and zero cents at valid discounts remains zero.
  • Add fractional-cent cases after agreeing the rounding rule; for nearest-cent rounding, (101, 50) === 51.
  • For [a:{cents:200}, b:{cents:100}], assert ascending result [b,a], original order still [a,b], and result array is a different reference. Repeat with a frozen input array to expose accidental mutation.
  • Cover empty, single-item and equal-price arrays. Check any explicit tie-order requirement if one exists.
  • Run the repository's existing type check and tests, then examine call sites for discount validation, pricing expectations and array ownership. No tests or call-site inspection were performed from this supplied diff; no security exploit or actual data loss is established.

Files from this run