I do AI quality control for a living: rubrics, golden answers, agent trajectories. These are the tools I built for my own paid jobs, given away. There is enough for everybody to eat.
Writes rows that are atomic, binary, self-contained and objective, and refuses to write the mushy kind.
Get it →
Audits a rubric you already have. Duplicates, inverted negatives, category imbalance. Feedback only.
Run it →
Grades your reference answer against your own rubric before you build a thousand tasks on it.
Try it →
A one page checklist that runs any criteria doc through the touring-rider standard. Works on specs too.
Grab it →
The platforms paying people to make AI less wrong. I work on several of them. Referral links stated plainly.
See the board →
The consulting side. Rubric architecture, model failure design, agent trajectory QC, creative tooling.
Next door →A Claude skill. Drop it in your skills folder and ask for a rubric. It writes rows that are atomic (one thing per row), binary (yes or no, no partial credit), self-contained (the grader never has to look anything up) and objective (two graders land on the same answer). If you ask it for a mushy row it says no and explains why.
What you get: the skill file, a worked example, and the three rules it refuses to break.
Paste in a rubric you already have. It looks for duplicated rows, negatives written backwards, categories that are overweighted, and rows that cannot be scored without guessing. It gives feedback only. It never rewrites your rubric for you, because that is your job and you know the task.
What you get: a plain-language report with the row number next to every flag.
A golden answer is the reference answer a task is graded against. Before you build a thousand tasks on one, run it against your own rubric. This tool scores the golden answer row by row and tells you where the reference itself would fail. It catches the embarrassing case where the answer key does not pass the test.
What you get: a scored table and a list of rows to fix before anybody else sees it.
A touring rider is the document that tells a venue exactly what an artist needs, down to the towel count. Good ones leave nothing to interpretation. This is a one page checklist that runs any criteria doc, spec or rubric through that standard. If a promoter could misread it, a grader will too.
What you get: the checklist as a printable page and a paste-ready prompt.
The platforms that pay people to make AI less wrong, listed plainly. I work on several of them. Where a link is a referral link I say so in the same sentence, and where a platform has a waitlist I say that too. Updated when something changes, not on a schedule.
What you get: the list, what each platform is looking for, and what the work is like day to day.
Sep 19Best Alternative to a Negotiated Agreement sounds like a minor Batman villain. It is your plan B, and it sets the floor for every deal you make.
Sep 16Six weeks of tweaking, then a month of leaving it alone. The instructions file, the skills, the MCP servers that stayed, and the ones I deleted without ceremony.
Sep 15Part editor, part teacher, part playtester. A plain-language tour of reading AI answers, grading them against a rubric, and explaining why the good ones are good.
Fuego!
Rubric architecture, model failure design, agent trajectory QC and creative tooling run through the studio, B1NARY H0RDE.
The Nonsense Emailer
What I tested, what I am reading, one tool that saved me an hour, and one thing that made me laugh. Sent when it is worth sending, never to hit a quota.
piddonkadonk