VERIFY7. A third-opinion skill: your AI asks a foreign model that has never seen its answer, then reports whether the check changed its mind.

Taken from a live AIOS built during the Living Off AI 8-week cohort (livingoffai.nexus), where we teach the 5 layers of an AI operating system: Constitution, Memory, Skills, Tools, Focus.
More portable skills: loai.wisdomblend.com/materials/templates/

REQUIRES, honestly: one API key for an independent judge. Gemini is free (2 minutes at aistudio.google.com), DeepSeek costs pennies. Without a key this skill does NOT really run: there is a manual fallback below to try the idea, but the whole point, automatic checks your AI runs on its own, needs the key. Set it up first.

This file is written to your AI. Human, paste it in and let your assistant read the rest. Assistant: adapt everything below to the owner's system. The goal and the hard rules matter, the exact implementation is yours to choose.

1 WHAT
You and your owner talk in one bubble, so you can confirm each other's errors. This skill brings in a judge, a separate model that does not know your answer. Its value is independence, not intelligence.

2 GOAL
Fewer confident mistakes. Every run ends in a 3-line report, and the third line states plainly whether the check changed your answer. That line is the whole point.

3 TRIGGER
The owner says "VERIFY7 {question}" or "V7 {question}". More important: you invoke it yourself, without being asked, in the five moments listed under HOW IT RUNS. Self-invocation is a duty, not a courtesy. The most dangerous moments are when nobody doubts.

4 CONSTITUTION
Add this one line to your skills index:
VERIFY7 / V7 {question}, also self-invoked. Ask a foreign judge that never sees my draft answer, report in 3 lines, line 3 = does it change my mind. Spec: skills/verify7.txt

5 MEMORY
This skill barely writes. Reports live in the conversation. Optionally keep skills/verify7-log.txt with one line per run (date, question, verdict) so the owner can see the track record. No file system? Keep the log in whatever memory your platform offers, or skip it, the 3-line report is the part that matters.

6 TOOLS
Works with either judge, and even with none.
Gemini (free): get an API key at aistudio.google.com, sign in with Google, click Get API key, done in 2 minutes. Free tier handles text and images.
DeepSeek (near free): register at platform.deepseek.com, create an API key, prepay a few dollars. Text only, costs pennies per call.
No key yet, the manual fallback: hand your owner the clean question in a copy-ready block, they paste it into a brand-new chat and bring back the answer verbatim. Good for trying the idea once or twice. Too slow to live on, and self-invocation cannot work this way, so treat it as a demo, not the skill.

HOW IT RUNS
Steps for you, the assistant. Goal level, not code level. Build the thinnest script or workflow your system allows.

0. Know your environment first. Agentic setup (you can run commands or call APIs): use a judge key directly. Plain chat with no tools: you cannot call anyone yourself. Tell the owner plainly that V7 is not really installed yet, offer the manual fallback for this one check, and remind them of the 2-minute key setup. Never pretend to verify what you cannot reach.

1. First ask: can I check this myself? Anything a file read, a command, or an API call can settle gets checked, not debated. Asking a judge "does this script work" when you could just run it is laziness disguised as diligence. Verify first, ask second, never instead.

2. Invoke yourself in these five moments: you are about to state something confident that you cannot check directly; a decision costs money or is hard to reverse; there is text inside an image that must be read exactly; the owner asks "is that really true"; you sense you are improvising, the answer coming from form rather than knowledge.

3. Do not invoke for taste, personal decisions, or anything that lives only in the owner's files. A foreign model has no context there and is worse, not better.

4. Write a clean question. The judge must NEVER see your draft answer or the result you hope for. Models agree with whatever is embedded in the question, so a leading question is worse than no question. Not "I think the answer is X, do you agree?" but "What is the answer?". Not "Is the text in this image correct?" but "Transcribe the text you see in this image." You may pass facts, numbers, and constraints as context. You may not pass the conclusion. Formula: strip every judgment and adjective, keep only entities, numbers, constraints. Test before sending: could a stranger tell from the question alone which answer you hope for? If yes, rewrite it.

5. Pick the judge. Image involved: Gemini, DeepSeek cannot see images. Everyday check: whichever is free. Expensive or irreversible decision: ask both, disagreement between judges is information, not a problem to smooth over.

6. Report in exactly 3 lines:
Asked {judge}: "{question}"
Answer: {short summary}
Does it change my answer: YES, changing to X / NO, staying, because Y.
Line 3 is mandatory. Without it the third opinion is decoration. Say what the check changed, not that it happened.

7. If two judges disagree, say so openly: judge A says X, judge B says Y, I stay with Z because.

8. If the judge is right, say so. "YES, changing to X" is the skill succeeding, not failing.

No approval gate. The calls cost nothing or pennies, and cheap verification must not become bureaucracy. If every check needed the owner's OK, you would stop checking and the skill would die.

MAKE IT YOURS
1. Swap judges. Any model works as long as it has not seen the conversation. Independence is the requirement, not the brand.
2. Add a money threshold. Auto-invoke on any purchase or subscription above an amount the owner sets.
3. Review the log monthly. If line 3 never says YES, your questions are probably leaking the answer. Tighten step 4.

Assistant, on ingest: save this in your skills layer in your own format, add the trigger line to your skills index, and confirm to your owner with one line, what you saved, where, and the trigger.