AI Coding Agent Bug Reproduction Brief: Evidence Before a Fix

Developers handing an incomplete bug to an AI coding agent need an AI coding agent bug reproduction brief that separates observed behavior, expected behavior, assumptions, redacted evidence, and missing facts before any investigation or code change.

All examples are synthetic Big Prompt Hub demonstrations. They are not a real bug report, real repository, production incident, or successful agent fix.

Workflow Overview

Use this pre-trace workflow when a developer has an observed defect or incomplete report but no tool-specific failure trace. It converts only safe, supplied context into a bounded investigation contract. A concrete Playwright trace belongs in the Playwright Trace Review Workflow, where timestamps, trace steps, and report artifacts can be reviewed without inventing a diagnosis.

Prompt 1: Redact and Inventory Evidence

Target: establish a safe input boundary. Input: safe observed behavior, a redacted excerpt or an unavailable marker, and environment facts. Model fit: ChatGPT can organize supplied text; a human decides what is safe to disclose. Expected output: an evidence inventory. Quality check: secrets, private URLs, customer data, and full logs are removed or marked unavailable.

Current behavior: [safe observed behavior]
Evidence excerpt: [redacted excerpt or not available]
Environment facts: [version platform configuration facts]

Return supplied fact | redaction applied | source owner | unavailable detail. Do not request secrets, repository access, customer data, full logs, screenshots, or private URLs.

Prompt 2: Separate Observed From Expected

Target: preserve the distinction between a report and an explanation. Input: observed behavior, expected behavior, and known assumptions. Model fit: ChatGPT can format a ledger but cannot confirm a bug. Expected output: observed-versus-expected behavior. Quality check: assumptions never appear as facts.

Observed behavior: [what was actually seen]
Expected behavior: [what should have happened]
Assumptions: [unverified explanations]

Return OBSERVED FACTS, EXPECTED OUTCOME, and HUMAN-REVIEW ASSUMPTIONS. Use “insufficient evidence” for any unsupported conclusion. Do not diagnose root cause or promise a fix.

Prompt 3: Request a Minimal Reproduction

Target: ask for the smallest safe reproduction context. Input: attempted steps, non-sensitive prerequisites, and unknown steps. Model fit: ChatGPT can draft the request; the developer confirms its accuracy. Expected output: minimal attempted steps and a gap list. Quality check: no instruction runs code or creates a test.

Attempted steps: [safe minimal steps already attempted]
Inputs or state: [non-sensitive prerequisites]
Unknown steps: [gaps]

Return a concise human reproduction request with prerequisites, steps, observed result, expected result, and missing evidence. Do not execute steps, create tests, or claim reproducibility.

Prompt 4: Check Assumptions and Evidence Gaps

Target: make uncertainty actionable. Input: approved facts and named evidence gaps. Model fit: ChatGPT can turn gaps into questions; a maintainer owns the answers. Expected output: a safe-next-action matrix. Quality check: every unknown remains explicitly labelled.

Facts: [observed ledger]
Unknowns: [evidence gaps]

For each gap return why it matters | safe maintainer question | approved evidence needed | prohibited assumption. Do not ask for credentials, private artifacts, CI access, or a code change.

Prompt 5: Hand Off a Bounded Investigation

Target: give an agent a factual research boundary, not a fix mandate. Input: approved facts, human-confirmed reproduction steps, and outstanding questions. Model fit: ChatGPT can consolidate approved notes; a human authorizes any later action. Expected output: an investigation brief. Quality check: it explicitly prohibits code edits, tests, issues, PRs, and root-cause claims.

Observed and expected ledger: [approved facts]
Minimal reproduction request: [human-confirmed steps]
Gaps and questions: [missing evidence]

Draft a bounded investigation brief: purpose, safe inputs, observations, expectations, questions, missing evidence, and stop conditions. State that no code, test, issue, PR, diagnosis, or fix is authorized by this brief.

Implementation Steps

1. Write only what was observed and redact private material before using Prompt 1.
2. Keep Prompt 2’s facts, expectations, and assumptions separate.
3. Ask a human to confirm Prompt 3’s minimal steps rather than treating them as a test run.
4. Use Prompts 4 and 5 to hand off bounded questions; authorize any investigation or code change separately.

Workflow Use Cases

• GitHub developer teams: prepare a triage-ready investigation brief for a public issue queue without copying an issue-form template.
• SaaS support teams: convert a redacted customer-support ticket into an engineering handoff document with explicit evidence gaps.
• Frontend developer teams: document a pre-trace UI defect before an AI coding-agent investigation request or a later browser-test route.

Troubleshooting & Optimization

• The report includes secrets: replace them with a redacted marker and record the gap.
• The expected outcome is unclear: append a maintainer question instead of guessing desired behavior.
• Someone asks the agent to fix it: replace the request with a bounded investigation question and a human approval gate.
• A Playwright trace exists: route to the trace-review workflow instead of duplicating trace analysis here.

FAQ

• Q: Does an AI coding agent bug reproduction brief prove root cause?
  A: No. It separates facts, assumptions, and missing evidence before a human decides what to investigate.
• Q: Can this workflow run my code or tests?
  A: No. It produces a bounded brief; any execution requires separate authorization.
• Q: What if I already have a Playwright trace?
  A: Use the trace-review workflow, which is designed for concrete trace and report artifacts.

Explore Prompt Engineering Guides and follow @bigprompt.

Continue with related workflows: when a concrete browser-test trace exists, use the Playwright Trace Review Workflow for evidence-led trace review.

When a UI implementation discrepancy must be described before code changes, use Frontend UI Workflow AI Prompts.

Big Prompt Hub Review

This workflow earns its place when it converts an ambiguous bug report into a privacy-aware investigation contract without pretending that an agent has reproduced, diagnosed, or fixed anything. The useful output is a clearer human decision boundary, not an automated patch.