The /raise-the-bar Skill

What it does

You finish something. Before it gets called done, a fresh critic that has never seen the conversation reads exactly two things: what you originally asked for, word for word, and what came out. It opens the real output, not a summary of it. Then it says SHIP or NOT YET, and if not yet, the single biggest gap. You fix that one thing. A new critic looks. Three rounds, max. If it's still not there, it says so plainly instead of rounding up to done.

That's the whole skill. It exists because the same mind that writes the work also grades it, and it grades its intentions instead of the output. A stranger only has the output.

When to reach for it

Anything you're about to hand to someone, where "did it cover what I asked?" has a real yes or no.

• An email that has to do a job. "Explain the price increase, keep the client, under 120 words." The critic's first pass on my test: NOT YET, it never gives her a reason to stay.
• A proposal, a quote, a plan. Scope, price, timeline, what's not included. Whichever one is missing is the gap.
• A report with numbers. If it says "revenue is up 12%" and nobody re-added the column, that's the biggest gap before anything about wording.
• A page or a form. "The signup works" is not a claim the critic accepts. It submits the form.
• Code. It reads the change and runs what's runnable.

Skip it for quick answers, opinions, chat, and anything you wanted fast and raw. It's for deliverables, not conversation.

Judge against the ask, or against an example

By default the bar is your request, read literally. But you can hand it something real to match or beat: your own best past work, a competitor's page, the deck a client loved last quarter. It asks once, "Is there something real this should match or beat?"

If you name one, it first writes a short checklist of what makes that example good, in plain words, and shows it to you. You fix anything it got wrong. Then every round is judged against that written checklist and the example side by side. That's what keeps the bar steady when the same thing gets made every month: the fourth one is held to the same standard as the first, and when the standard changes, one file changes.

The honest part

This is adapted from Matt Shumer's gauntlet loop: a builder, a separate harsh critic, loop until it passes. His version compares against a real reference (his demo was a game judged against actual Call of Duty). Mine adds two things. The critic can judge against the original ask when there's no reference to compare to, which is most everyday work, and the loop has a hard stop so it can't run forever inventing its own standard. No text was copied.

It is also the last stage of hatch, pulled out as one file. Hatch is for when you don't know what to build yet. This is for when you built it and want to know if it's done. Most days I use this ten times and hatch once.

It's working if

• The first verdict is NOT YET more often than SHIP. A critic that always ships isn't judging.
• The gap it names is one you'd have missed, and you agree with it once you read it.
• Round two usually ships. If you're hitting round three every time, your asks are too vague; fix the ask.
• You can hand the thing over without re-reading it yourself.

Common questions

Isn't this just asking the AI to review itself?

No, and that's the point. Reviewing itself inline is the thing this kills. The critic is a separate fresh context that never saw the work being made.

Why only one gap per round?

A list of ten notes turns into ten cosmetic edits. One big fix changes whether the thing gets accepted, and the small gaps usually dissolve after it.

Why no score out of 10?

Scores drift. 7 becomes 8 becomes "good enough." SHIP or NOT YET can't drift.

Does it need Claude Code?

Yes, for now. The fresh critic is a separate agent, and that's what a chat window can't give you. A paste-in version may come later.

What does it cost?

Free, MIT. Each round is one extra AI call. Three rounds is the cap.

Where it fits

Second skill in the Think group. Hatch is the front half: decide what to build. This is the back half: don't call it done until a stranger agrees. Every skill I release after this one ends by running it.