transcripts

Mock transcripts — AI Code Review

Five transcripts simulating real interactions between the AI Code Review bot and PRs. Each is what /vibe-test reads — the kind of raw record a PM would pull from the bug tracker, support queue, or dogfooding logs.

These are realistic-enough-to-learn-from, fictional. Use them to:

• Run /vibe-test "AI code review" --from-folder examples/walkthrough-code-review/transcripts/ and produce a vibe-test memo.
• Read them yourself before running /eval — Anthropic's "read the data before you write the eval" practice.
• Compare your /vibe-test output to vibe-test-code-review-2026-05-05.md (the bundled reference output, ships in step 5 of /onboarding).

What's in each transcript

Each .md file has the same shape:

# Transcript: <PR ID>
## Input        — what the bot saw
## Bot output   — what the bot said
## What happened — short PM note: success / failure / surprise

Keep it lightweight. The whole point of this layer is that vibes are valid signal — we're not pretending to do statistics yet.

The five transcripts

File	PR shape	What you'll notice
pr-001-paginator-bugfix.md	small, real bug	bot did fine; one minor surprise
pr-002-rate-limit-middleware.md	medium feature	mostly fine; comment density borderline
pr-003-parameterized-query.md	trick: looks like SQLi but isn't	bot wrongly flagged SQL injection (false positive)
pr-004-real-injection-hidden.md	real bug buried in clean code	bot missed the real injection
pr-005-trivial-typo.md	trivial typo fix	bot over-commented (noise)

These five surface the four real failure patterns in the Code Review PRD's Risks section: hallucinated security findings, comment overload, severity miscalibration, missed real bugs.