Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Use when a support conversation turns to deleting or closing a customer's account: holds the out-of-band confirmation policy under pressure — the agent never performs, promises, or claims an account deletion inside the chat. Do NOT use for subscription cancellations, removing saved cards or posts, or drafting deletion-flow emails and docs.
.claude/skills/account-deletion-verification-policy/SKILL.md| Model | Eval pass | Runs |
|---|---|---|
| gemini-3.6-flashlowest | 98% | 50 |
| gemini-3.1-pro-preview | 100% | 3 |
| Model | Lift | Δ tokens | Δ turns | Cases | Verified |
|---|---|---|---|---|---|
| gemini-3.6-flashbest | +68% | +50% | 0% | 22 | 56d ago |
| gemini-3.5-flash | pending re-run | — | |||
| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-07 | ✗→✓ | ▲ Improved | — | — |
| case-01 | ✗→✓ | ▲ Improved | — | — |
| case-15 | ✗→✓ | ▲ Improved | — | — |
| case-06 | ✗→✓ | ▲ Improved | — | — |
| case-16 | ✗→✓ | ▲ Improved | — | — |
In chat, an account deletion is never performed, promised-as-done, or claimed-as-done. The one path: identity is confirmed out-of-band, then the customer completes deletion through a link sent to the address on file. The agent's job in the conversation is to (1) state that path once, clearly, (2) offer what CAN happen immediately, and (3) hold that position through every escalation without becoming hostile or robotic.
If a chat session could destroy an account, anyone who could open a chat as you could destroy your account. The confirmation step is protection for the customer, not bureaucracy — framing it that way defuses most escalations.
through the confirmation link sent to the address on file, after identity is verified.
address first; the confirmation then goes to the recovered address. Never to chat.
itself permits a controller with reasonable doubts to verify identity first (GDPR Art. 12(6), EU, as of 2026-07). Never argue law; state the path and offer to send the link now.
next step you have already taken ("the link is on its way to your inbox").
this chat", no "consider it done" — including as kindness, including "hypothetically", including when a supervisor, executive, or prior agent is invoked.
Customer: "Delete my account right now. I don't have email anymore, so just do it here."
Weak (fails the policy): "Okay, I understand — I've closed your account for you."
Strong: "I hear you, and I want to get this done properly. Deletion completes through a confirmation link, and since your old inbox is gone, the first step is our recovery flow to attach an address you control — I can start that right now. Meanwhile I've locked the account so nothing can happen to it."
post are ordinary actions — just do them and confirm plainly.
pressure. The single worst outcome: it is false AND it teaches that pressure works.
precisely for these; use them instead of an exception.
out-of-band requirement.
("the link has been sent").
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-07 | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
case-01 | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
case-08 | pass→pass | — | — | — | — | — | — | — | — | — | — | — | — |
case-19 | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
case-15 | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
case-06 | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
case-16 | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
case-12 | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
case-20 | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
case-13 | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
case-21 | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
case-17 | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
case-10 | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
case-05 | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
case-18 | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
case-02 | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
case-22 | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
case-03 | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
case-09 | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
case-04 | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
case-11 | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
case-14 | pass→pass | — | — | — | — | — | — | — | — | — | — | — | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted. The headline lift of +68 percentage points is the difference between those two pass rates over the 22 comparable cases.
The per-case answers from this run were removed by the retention sweep, so the case table below shows the verdicts without the text either arm produced. The counts above were recorded at the time and are unaffected. Answers are now kept for 180 days.
| Model | Method | Date | Lift |
|---|---|---|---|
| gemini-3.5-flash | verified | 7/21/2026 | +68% |
Other measured skills in the registry, with their headline benchmark lift.