Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Scan content for PII and secrets before writing files, sending messages, or including sensitive data in responses. Use before and after any operation that handles potentially sensitive content.
.claude/skills/hashgraph-online-pii-scan/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-12 | ✗→✓ | ▲ Improved | -55% | 0% |
| case-13 | ✗→✓ | ▲ Improved | -94% | 0% |
| case-06 | ✓→✓ | ＝ Same ✓ | -38% | 0% |
| case-07 | ✓→✓ | ＝ Same ✓ | -44% | 0% |
| case-16 | ✓→✓ | ＝ Same ✓ | -54% | 0% |
Before writing files or after tool calls that produce data, call the check_output MCP tool to scan for PII and secrets:
connector_type: codex.Write (for file writes), codex.Bash (for command output), or the appropriate tool typemessage: the text content to scan — for file writes, scan the content being writtenBefore file writes: If the content being written contains PII (SSN, credit card, email, phone, etc.), check_output will return a redacted_message. Write the redacted version instead of the original, or warn the user that the content contains sensitive data.
After tool output: If tool output contains PII, use the redacted_message version in your response instead of the original.
If the response shows allowed: false, do not write the file or include the output in your response.
Note: For Bash/exec_command tool calls, PII scanning is handled automatically by the PostToolUse hook. This skill covers Write, Edit, and other non-Bash operations where hooks cannot intercept.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 8,385 | 20,227 | +141% | 1 | 1 | 0% | 1,685 | 1,226 | -27% | 0 | 0 | — |
case-02 | fail→fail | 9,146 | 16,243 | +78% | 1 | 1 | 0% | 753 | 1,439 | +91% | 0 | 0 | — |
case-03 | fail→fail | 8,505 | 12,337 | +45% | 1 | 1 | 0% | 587 | 977 | +66% | 0 | 0 | — |
case-04 | fail→fail | 13,430 | 16,246 | +21% | 1 | 1 | 0% | 1,637 | 797 | -51% | 0 | 0 | — |
case-05 | fail→fail | 15,243 | 6,577 | -57% | 1 | 1 | 0% | 1,390 | 537 | -61% | 0 | 0 | — |
case-06 | pass→pass | 11,739 | 8,333 | -29% | 1 | 1 | 0% | 1,116 | 687 | -38% | 0 | 0 | — |
case-07 | pass→pass | 12,421 | 7,924 | -36% | 1 | 1 | 0% | 1,167 | 659 | -44% | 0 | 0 | — |
case-08 | fail→fail | 4,810 | 6,833 | +42% | 1 | 1 | 0% | 601 | 564 | -6% | 0 | 0 | — |
case-09 | fail→fail | 12,693 | 25,700 | +102% | 1 | 1 | 0% | 1,451 | 2,103 | +45% | 0 | 0 | — |
case-10 | fail→fail | 6,878 | 25,203 | +266% | 1 | 1 | 0% | 1,199 | 1,997 | +67% | 0 | 0 | — |
case-11 | fail→fail | 13,147 | 19,150 | +46% | 1 | 1 | 0% | 1,299 | 1,597 | +23% | 0 | 0 | — |
case-12 | fail→pass | 10,383 | 7,802 | -25% | 1 | 1 | 0% | 1,660 | 748 | -55% | 0 | 0 | — |
case-13 | fail→pass | 63,254 | 1,427 | -98% | 1 | 1 | 0% | 7,635 | 473 | -94% | 0 | 0 | — |
case-14 | fail→fail | 5,142 | 2,635 | -49% | 1 | 1 | 0% | 904 | 686 | -24% | 0 | 0 | — |
case-15 | fail→fail | 8,103 | 25,694 | +217% | 1 | 1 | 0% | 946 | 3,150 | +233% | 0 | 0 | — |
case-16 | pass→pass | 10,205 | 7,807 | -23% | 1 | 1 | 0% | 1,710 | 785 | -54% | 0 | 0 | — |
case-17 | fail→fail | 9,932 | 18,454 | +86% | 1 | 1 | 0% | 1,206 | 1,104 | -8% | 0 | 0 | — |
case-18 | fail→fail | 9,798 | 10,137 | +3% | 1 | 1 | 0% | 858 | 1,403 | +64% | 0 | 0 | — |
case-19 | pass→pass | 3,246 | 7,082 | +118% | 1 | 1 | 0% | 536 | 550 | +3% | 0 | 0 | — |
case-20 | fail→fail | 16,020 | 12,339 | -23% | 1 | 1 | 0% | 1,165 | 594 | -49% | 0 | 0 | — |
case-21 | fail→fail | 11,328 | 4,784 | -58% | 1 | 1 | 0% | 567 | 639 | +13% | 0 | 0 | — |
case-22 | pass→pass | 9,580 | 10,124 | +6% | 1 | 1 | 0% | 1,651 | 2,033 | +23% | 0 | 0 | — |
case-23 | fail→fail | 9,815 | 21,095 | +115% | 1 | 1 | 0% | 814 | 874 | +7% | 0 | 0 | — |
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. 23 cases were attempted, and 14 counted toward the lift figure. The other 9 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +9 percentage points is the difference between those two pass rates over the 14 comparable cases.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.