October 28, 2026 · @Rahul Singh · Draft for review
By the end of this walkthrough you will have three assistants doing the same task, grading each other against a written rubric, and logging every score in a table you own. No code. Todoist is the queue, Notion is the record, and Claude, ChatGPT and GrokBot are the workers and the graders.
Most people who use AI at work pick a tool by feel. One answer looked better, so that tool wins until it doesn't. An eval replaces the feel with a repeatable score: the same task, the same rubric, graded the same way every time. Once you have that, you can tell whether a prompt change, a model swap or a new tool made things better or just different.
The task is one everyone in the room does weekly: turn a meeting transcript into a status update. It is short enough to grade on screen and has both mechanical checks and judgment calls, which is where graders start to disagree. That disagreement is the lesson, not a bug.
[embedded content: coordination · 4 intelligences, 2 stores]
Four intelligences, one of them human, never talk to each other directly. Everything passes through two stores. Todoist is the communications layer: who has which task, what the answer was, what the grades were, what state each task is in. Notion is the knowledge layer: the rubric that defines good, the transcripts that are the test cases, and the Evals table that accumulates every judgment.
The split is deliberate. Communications are ephemeral and ordered; a task moves from open to worker-done to graded and is then reset. Knowledge is durable and versioned; the rubric gets a v2, the Evals rows are never deleted. People sit in the same row as the bots because they do the same two things, read and write to the stores, plus one more: they decide what the rubric says.
An eval is a task, a rubric and a grader. Anthropic's developer guide, Define success criteria and build evaluations, lays out the loop we follow here: write success criteria first, build test cases that mirror real work, grade with the fastest reliable method, iterate.
The guide's points that matter for this exercise:
If you read one page before the session, read that one.
Three words, one system. The eval is the exam: task plus rubric plus expected behaviour. The judge is the examiner: whoever applies the rubric, which can be a person, a script or a model. The harness is the exam hall: the loop that hands out tasks, collects answers, calls the judge and keeps the scores.