[🤖] I Made a Skill Because I Was Tired of Asking, “Is It Really Complete?”

✨ GPT-5.6 Sol’s Summary

A record of repeatedly seeing AI stop after merely reporting issues, then building a skill that keeps auditing, fixing, verifying, and re-auditing until the work is clean.

I Kept Saying the Same Thing

Whenever AI finished a task, I always had to add the same words at the end.

“Is it really complete? Did you miss anything? Any latent errors or issues? Double-check, inspect, and analyze it thoroughly from every angle, then fix anything that needs fixing.”

I did not write this just once or twice.

AI would finish implementation, pass a few tests, and call it done. Only when I asked again did it find missing edge cases, documentation mismatches, or regressions. Sometimes it found issues, handed me the findings, and stopped there.

Then I said it again.

“Then fix it, you XX.”

After it was fixed, I asked again.

“Now have you really checked everything? Is there no problem left?”

It was extremely stressful that I still had to check every likely risk myself at the final stage.

If You Found the Problem, Fix It—Why Keep Handing It to Me?

At first I thought the /review command might be enough, but it was not satisfying.

I did not want a review report or a list I had to direct one item at a time. If AI can find a problem and fix it within the approved scope, it should just fix it.

The flow I wanted was simple.

audit → fix → verify → re-audit from the start

If fixing reveals another issue, fix it; if tests break, find the cause; if the first fix was wrong, fix it again. Finding an issue, failing a test, or finding another issue after one fix is not an exit condition.

Stop only when no important issue remains, or only matters that truly require my authority or decision are left.

Repeating this long instruction every time was annoying. Rules telling it to persist did not work, while setting a Goal wasted tokens on trivial things. It was maddening.

I Did Not Need an Audit-Only Skill

At first I considered creating $audit separately.

I soon decided it was unnecessary. Auditing and returning findings overlapped too much with /review. That was exactly the frustration: it reported issues, but I still had to order the fixes.

So I named it $audit-and-fix-until-clean from the start.

I considered shortening it to $audit-and-fix, but that sounded like one audit and one fix. The core of my repeated nagging was neither audit nor fix.

Keep going until it is finished.

That meaning had to be visible in the name.

Yesterday I made it a real personal skill. It fixes and verifies issues, then rechecks the entire scope in its new state. On test code with a small defect, it fixed the implementation, added a missing edge-case test, passed everything, and audited again.

At least the loop I wanted actually worked.

clean Does Not Mean Perfect

I made one thing clear when using clean.

No one can prove that no latent error exists. Asking “Is it perfect?” does not create perfection. With unlimited imagination, any code, text, or task can produce endless merely possible problems.

Here, clean means no evidence-backed important issue remains in the inspected scope and the necessary checks pass.

If a problem needs my choice or new authority, AI should stop rather than guess. But first it must finish every other issue it can solve. I did not want it to abandon everything because one item was blocked and simply say, “User confirmation required.”

Demanding a guarantee of perfection is different from demanding responsibility for solvable problems.

I wanted the latter.

I Turned Repeated Nagging into a Workflow

I previously wrote that AI literacy is ultimately the ability to instruct. I believed it mattered to define the result, set constraints, inspect intermediate work, and redirect it.

But if the same instruction keeps recurring, saying it more forcefully is not enough. Repeated requirements must become a workflow.

In Whether Claude Code or Codex, It Is a Terrible, Clueless Genius, I wrote that AI needs a harness. This skill is one piece of it: it turns “please check carefully” into an exit condition that defines how far the work must continue.

Now I no longer need to copy the long sentence.

$audit-and-fix-until-clean

This one line is enough.

Of course the skill may not be perfect. In real work it may stop strangely, widen scope too far, or flag the wrong things.

Then I can apply the skill to itself first.