AI Feature Pre-Ship Checklist

---
name: ai-feature-pre-ship-checklist
description: Use before shipping an AI-built feature to production users. Produces a verification checklist scoped to the feature, focused on AI-specific failure modes. Triggers on "before I ship", "pre-deploy check", or "AI built this, is it safe".
---

# AI Feature Pre-Ship Checklist

AI-built features fail in patterns. The hallucinated API call that worked in the happy path but breaks on the second customer. The silent error swallow that hid a real bug. The validation the AI removed because it "didn't see why it was needed." Standard pre-deploy checklists don't catch these.

This skill produces a feature-scoped checklist focused on the AI-specific failure modes.

## Required inputs

1. **What the feature does** — in 1-3 sentences, plus who uses it
2. **What it touches**:
   - Money / billing? (Stripe, invoices, subscriptions)
   - User data? (PII, uploads, exports)
   - External APIs? (which ones, what writes)
   - Database mutations? (which tables)
   - Permissions / auth?
3. **How much of it AI wrote** — all of it, most of it, the framework but I customized
4. **Which model / tool** wrote it (Cursor, Claude Code, ChatGPT, v0, etc.)

If the user can't answer #2, ask before producing the checklist. The blast radius determines what to check.

## Checklist generation

Build the checklist in five sections. Drop sections that don't apply.

### 1. Hallucination check
- [ ] Every external API call exists in the actual SDK / docs at our version
- [ ] Every database column / table referenced exists in the current schema
- [ ] Every internal function called exists and has the signature the AI used
- [ ] Run the feature with logs on; watch for "method not found" / "column not found" errors

### 2. Behavior preservation (if the feature touches existing code)
- [ ] Diff the AI-modified files; for each modification, confirm the change was intended
- [ ] Specifically watch for: removed validation, swallowed errors, renamed variables, changed defaults
- [ ] Re-test paths the AI didn't change but might have affected

### 3. Edge-case check (the cases the AI didn't think of)
- [ ] Empty inputs (empty string, empty array, null, missing field)
- [ ] Very large inputs (file size limits, list size limits)
- [ ] Wrong-type inputs (string where number expected, etc.)
- [ ] Concurrent / race cases (two users hitting the same endpoint)
- [ ] Permission boundary (user A trying to access user B's data)
- [ ] Idempotency (what happens if the request is retried?)

### 4. Money / data safety (if applicable)
- [ ] Every write to billing has a corresponding read-back / verification
- [ ] Refunds and cancellations work, not just charges
- [ ] User-uploaded files are size-limited and content-type-checked
- [ ] User-input is escaped at every output surface (HTML, SQL, shell)
- [ ] Auth checks are at the route level, not just UI level

### 5. Operational readiness
- [ ] Logging captures enough to debug a customer report ("user X says feature Y didn't work")
- [ ] Errors are surfaced (not swallowed); a real human will see them
- [ ] Rate limits exist if external API calls cost money
- [ ] There's a way to turn the feature off (feature flag, env var) without redeploying
- [ ] Someone other than the builder has clicked through the happy path

## Output format

```
## Pre-ship checklist: [feature name]

Scope: [what the feature does, blast radius]

### Hallucination check
- [ ] ...

### Behavior preservation
- [ ] ...

[etc.]

## High-risk items (do these first)
1. [the items most likely to break for this specific feature]

## What to skip if time is tight
[items that are nice-to-have for this feature's blast radius]
```

## Anti-patterns

- Generic "write tests" advice — the AI may have already written tests for the wrong cases
- Skipping the diff review because "all tests pass" — AI tests test what the AI thinks the code does
- Trusting the same model to verify its own output — it has the same blind spots

## Tone

- Concrete. "Refunds and cancellations" beats "test edge cases."
- Calibrated to blast radius. Money/data features get more checks; internal-only tools get fewer.
- Honest about what AI-built code skips by default. The checklist exists *because* the AI didn't think of these.