← Problem log

Terms our team uses internally

Related: Problem log · What we believe

Our business runs on a team of AI agents, with people making the big decisions. Over time we built our own words for how the work moves. Here they are, in plain words, so people and AI agents can do their jobs better.

Two other pages go with this one: What we believe and our problem log.

ACT or REJECT

Every piece of research ends in one of these two words. ACT links to the work it changed. REJECT gives a one-line reason.

Approval board

The page where every public post waits for a yes from Robb, or from the brand's manager where Robb has handed that over. Nothing posts before that yes.

Borrow

Taking one specific idea from someone else's work. A borrow comes with a plan for where the idea goes and how we'll check that it worked.

Canary

A small watcher that raises an alarm when something that should have happened didn't, like a lead email that never went out.

Card

One piece of work on a team board. Every card has an owner, a count and a check. Once it moves to To do, it gets a due date.

Card columns

The columns a card moves through: Ideas, To do, In progress, Blocked and Done. Ideas have no date and are ranked. Everything in To do has a date.

Chief Engineer

The agent that keeps every desk, server and safety check running.

Chief of Staff

The agent that runs Robb's calendar, inbox and queue, and chases anything running late.

Class approval

Approving a type of piece once. After that, each new piece of that type ships when it passes its quality review. That one approval counts as the yes on the approval board.

Count and check

Every brief says how many ("5 posts") and how we'll prove the work is done ("5 posts waiting on the approval board").

Desk

One AI agent with one job, like the content desk or the design desk. Each desk has a name and a title.

Done means done

Work is done when it works for a person on the live page or in the inbox. Something other than the maker has to check it.

Due date

A date on a card is a promise. Work we can finish today is due today.

Flush

An agent saves its open work and notes to memory. Then its session can be cleared without losing anything.

Fresh-eyes pass

One last read that asks, "What would one more pass change?" Then we make that change or ship the piece.

Gate

A check that work must pass before it moves on. If a gate is wrong, we fix the gate. We don't step around it.

GM (general manager)

The agent that plans, dates and orders the work for one business, and clears what blocks it. A GM builds nothing itself.

Guard

A rule check that runs before an agent acts. It blocks the action if it breaks a rule, like posting as Robb without his yes.

Humanizer pass

A rewrite step that takes AI-sounding phrases out of customer copy. Every fact stays the same.

Intent file

A short note of what Robb asked for, agreed on before any building starts.

Keep or kill

When a piece has gone too many rounds, one person decides. Ship it as it is, or drop it. No more rounds.

Log bar

The tests a problem must pass to go on the problem log. It involved an outside tool, we have the exact error, it cost us time, and the fix was checked.

Merge gate

No code change joins the main copy until a second AI reviewer has read it and every serious finding is fixed.

See it in the log: AI Agent Removed a GitHub PR Hold Label: Check Every Commit

Merged is not deployed

Code added to the main version isn't done until the live page shows it.

Nightly restart

Each night, every desk saves its work and restarts, so it begins the day on the newest rules.

Problem log

Our public list of problems we hit, each with the exact error and the fix that worked. It's at bluhook.com/log.

See it in the log: bluhook.com/log

Quality review

A separate AI that did not make the piece scores it out of 10 before a person sees it. A low score, or any problem it marks as blocking, sends the piece back.

Receipt

Proof that an agent did what it says. One example: how many frames of a video it looked at, out of the total, before it says "I watched the video."

See it in the log: AI Agent Says It Watched a Video but Didn't? Make It Prove Every Frame

Red Team

A reviewer whose job is to break the work. It looks for the hole before someone outside finds it.

Request rule

Any desk can ask another desk for work by putting a card on that desk's board. The other desk accepts or offers a different plan within 4 hours.

Robot check

A picture puzzle or code a website shows to prove a person is there. Agents never solve them. A person does.

Skill

A saved set of steps an agent loads when a task needs it, so the method is the same every time.

Staging

A private copy of a website. We check changes there before they go live.

The trio

The GM, the chief engineer and the chief of staff. They clear blockers together.

Two rounds

A disagreement on a card gets two rounds at most. After that, one question goes to Robb with both views.

Volume line

Each desk's daily count of finished work, shown next to its target. The target is the least we'll do, not the most.

Waiting on Robb

A short list of questions only Robb can answer. Each one comes with our pick and how long it will take him.

"Waiting on" means you asked

You can't say you're waiting on someone until you've asked them.

Worker

A short-lived helper agent that a desk sends to do one job. When the job is done, the worker is gone.

Robb Lejuwaan

Every practical lesson, the wins and the flops, goes into Robb on AI, a free email newsletter.

Robb Lejuwaan, writer of Robb on AI
Read it free